Building Effective AI Agents:

Architecture Patterns and

Implementation Frameworks
Accelerate your enterprise AI transformation
with proven strategies from Anthropic's
customers and internal teams.
Contents
Executive summary 4

Common use cases and applications for AI agents 7

Common architecture patterns 10

Looking forward: the future of building AI agents 27

Next steps 29

2
Chapter 1

Executive summary 3
 Chapter 1

Executive summary

Generative AI answers questions. AI agents solve AI agents open up countless possibilities for organizations of every size and
problems. sector, but implementing them requires careful consideration of architecture
patterns, cost management, and operational governance.
For companies across industries, agents offer the potential for scaling operations
in ways current automation could never deliver: open-ended problem-solving,
dynamic decision-making, and complex multistep processes where the path The business case for AI agents
forward isn't predetermined.
Think of an AI agent as a smart digital assistant that can work independently
Organizations are seeing significant results from autonomous agents in to solve complex business problems by using tools that connect to your real
production. For example, Coinbase, the leading cryptocurrency exchange systems. At its core, an AI agent represents a sophisticated evolution of large
managing $226 billion in quarterly trading volume, built agentic customer language models that can autonomously direct their own processes and tool
support systems powered by Claude. Their Claude-powered agents handle usage to accomplish complex tasks.
thousands of messages per hour while maintaining 99.99% availability—critical
Traditional automation requires rigid prewritten scripts with every step mapped
when customers need constant access to their funds. The platform has spawned
in advance. Agents work differently. They assess a task, choose appropriate tools,
35-50 internal AI applications, transforming how the company serves millions of
try approaches, evaluate results, and adjust strategies as needed, much like how
users globally.
a skilled employee tackles unfamiliar projects. For example, an agent handling
Tines, the workflow orchestration and automation platform for security and IT customer support escalations could read the issue, check account history, consult
teams, built agentic workflow systems with Claude. Their agents dynamically knowledge bases, draft personalized responses, and loop in specialists, all
handle workflow logic during execution, collapsing complex multi-step security without human intervention.
operations into single-agent operations, corresponding to 100x time-to-value
What makes these systems powerful is their capacity for autonomous reasoning
improvement.
and tool selection, combined with the ability to recover from errors and maintain
And Gradient Labs, the company building customer operations agents persistence toward goal completion. Unlike traditional workflows where
for financial services, deployed a customer support agent with Claude that predefined code paths orchestrate AI interactions, agents maintain dynamic
understands customer queries within context and executes standard operating control over their decision-making processes, adapting based on environmental
procedures. Achieving 80-90% resolution rates, their agents can handle complex feedback and intermediate results.
workloads with limited human intervention, enabling employees to focus on
relationship building and other strategic work.

4
This makes them particularly valuable for scaling complex operations where
exact steps can't be predetermined, such as incident response, data analysis,
customer onboarding flows, or development workflows where automated testing
creates feedback loops for iterative problem-solving.

What organizations are achieving

The organizations deploying agents are seeing numbers that matter.

At a retail bank, for example, AI agents transformed credit risk memo creation.
What used to take relationship managers weeks of manually reviewing ten
different data sources now delivers 20 to 60 percent productivity gains and cuts
credit turnaround time by 30 percent. A European equipment manufacturer with
over €10 billion in revenue mapped out their agentic AI strategy and found

• from teams running these systems in production

• Architecture patterns from single agents to multi-agent orchestration, with

clear guidance on matching your specific problem to the right pattern(s)

• Technical requirements including API capabilities, tool integration, and

memory management for production-ready agents that actually work at scale

• Security and compliance frameworks for protecting sensitive data while

managing the unique risks that come with autonomous systems

• Implementation strategies for building teams and infrastructure that scale

with model improvements (instead of fighting against them)

• Future-readiness indicators to help you build systems that grow more

powerful as the underlying models improve – without growing more complex

Let’s dive in.

5
Chapter 2

Common use cases and

applications for AI agents 6
 Chapter 2

Common use cases and applications forAI agents

To understand where agents deliver real business value, let's examine how across over 25,000 customers. The platform delivers a 51% average resolution
organizations are deploying them today. Real-world implementations reveal rate out of the box (before customization), reduces response times from 30
patterns that can guide your own strategic decisions and help you identify the minutes to seconds, and supports over 45 languages.
highest-impact opportunities within your organization.
AI-enhanced human support coordination: Assembled uses Claude to power
their Assist platform, achieving a 20% increase in customer satisfaction while
Coding decreasing support spend, over 50% reduction in escalations, and more than
30% improvement in cases solved per hour. Their approach focuses on resolving
Accelerated development on enterprise systems: Augment Code uses Claude
complex Tier 2+ issues rather than just deflecting simple queries.
on Google Cloud's Vertex AI to help developers navigate complex codebases
with millions of interdependent lines of code. One enterprise customer
completed a project in 2 weeks that their CTO estimated would take 4-8 months, Legal
while developer onboarding accelerated from weeks to 1-2 days. The platform
Enterprise legal knowledge at scale: Thomson Reuters uses Claude in Amazon
helps teams understand sophisticated software systems faster, enabling them to
Bedrock to power CoCounsel, delivering expertise from 3,000+ subject matter
write, document, and maintain code more efficiently..
experts and over 150 years of authoritative content to legal and tax professionals.
The platform processes complex contracts and tax documents with rigorous
Data analysis accuracy through expert validation, with customers reporting they "can easily
see it cutting the time in half, maybe even more" and describing CoCounsel's
Conversational observability data exploration: Grafana uses Claude to power
efficiency as "staggering" - freeing professionals to "focus on higher-level, more
an intelligent assistant that enables teams at all skill levels, from CTOs to junior
strategic work."
engineers, to unlock observability data through natural language. Users can
ask questions like "What's the request latency for my checkout service?" and Flexible legal AI with superior instruction-following: Legora uses Claude to
Claude automatically finds relevant metrics and constructs appropriate PromQL power their entire legal platform, achieving 18% higher performance on their
and LogQL queries. proprietary large legal evaluation set across complex tasks. Claude Sonnet’s
gains come from "consistency over large tasks and documents and accurately
following complex instructions," enabling Legora to build flexible agentic
Customer support and operations
workflows that adapt to different practice areas and client requirements, helping
High-resolution automated support at scale: Intercom's Fin AI agent, powered lawyers "review and research with precision, draft smarter, and collaborate
by Claude, achieves up to 86% resolution rates with human-quality responses seamlessly."

7
Marketing
Automated multi-platform advertising at scale: Advolve uses Claude to
orchestrate their entire digital customer acquisition process, managing millions
of ads simultaneously across platforms with real-time data validation and
dynamic budget allocation. The system achieves a 90% reduction in operational
work time and a 15% increase in customer return on ad spend (ROAS), with
their platform reaching "human-level ROAS when managing multi-million dollar
budgets in under 30 days" for major enterprise clients managing ad budgets
exceeding $100M.

Vertical spotlight: Financial services

Automated fraud detection and risk assessment: Inscribe uses Claude to
power AI Risk Agents that reduce fraud review time by 20x—from 30 minutes to
90 seconds—while increasing output by 70x in client examples. The AI Fraud
Analyst detects fraud in images and PDFs, verifies applicant details through KYC
and KYB checks, uncovers risky transactions, and provides auditable risk reports
in approximately 90 seconds. This enables financial institutions to expand access
to creditworthy but underserved populations including thin-file, unbanked, and
credit-invisible individuals.

8
Chapter 3

Common
architecture patterns 9
 Chapter 3

Common architecture patterns

These use cases demonstrate significant agent potential across industries, but processing thousands of straightforward customer support tickets or extracting
successful implementation also depends entirely on choosing the right models, data from forms, a lighter, faster model will get the job done just as well at a
technologies, and an architectural approach. A customer support agent handling fraction of the cost.
routine queries needs a fundamentally different design than a multi-domain
The performance spectrum ranges from models optimized for the most complex
research system analyzing complex datasets.
reasoning tasks down to models designed for high-volume, straightforward
Understanding these architectural patterns helps you match technical applications. This means you can match your specific use case to the appropriate
complexity to business requirements while avoiding over-engineering that level of capability. Running a simple task through a premium model isn't just
increases costs without delivering proportional value. wasteful, it's also slower and more expensive at scale. When you're processing
hundreds or thousands of requests, those differences compound quickly.
Let's begin with the foundational design principles that guide successful AI
agent implementations. Practice modular design. This space moves quickly; new capabilities and
features emerge regularly. Design your systems for modularity so you can
evolve your agent's capabilities without needing to radically redesign your
Agent design best practices
infrastructure. Your architecture should bend with progress, not break under it.
Start simple, scale intelligently. As discussed in Building Effective AI Agents,
• Agent modularity, for example, might leverage a composition pattern where:
we suggest teams begin with single-purpose agents that do one thing well, then
gradually develop them into more sophisticated systems as your requirements • Prompts are defined in centralized configuration files or libraries
evolve. Simple systems are cheaper to run (fewer tokens, less compute), easier
• Tools as discrete reusable modules
to debug when things go wrong, and give you clear metrics that actually tie to
business outcomes. Agents are defined as needed, leveraging only the tools and/or resources needed
to accomplish their assigned task.
Choose the right model for the job. There are a lot of AI models out there with a
lot of different abilities and choosing the right one is critical. This composition pattern might have separate modules for web search, database
queries, and email composition, along with prompt templates for different
The key is balancing three factors: capabilities, speed, and cost. Think of it like
reasoning styles (analytical, creative, technical) or for different roles. Following
choosing the right tool from a toolbox: you wouldn't use a sledgehammer to
it lets you quickly define necessary agents during development and select
hang a picture frame, and you wouldn't use a tack hammer to demolish a wall.
predefined resources as needed.
For instance, if you're building multi-agent coding frameworks or doing complex
financial analysis, you'll want the most capable model available. But if you're Modular agents like these can scale organically, particularly those built on

10
frameworks like LangGraph or Mastra. As new AI capabilities emerge, a Build observable systems that explain themselves. AI applications often
componentized agent architecture offers natural integration points without function like black boxes, and agents add additional layers of complexity.
system-wide refactoring. You can easily integrate new tools into modular Beyond standard practices like structured logging, centralized monitoring, and
agent frameworks, and update the central configuration to roll out enhanced distributed tracing, AI applications introduce unique observability challenges
prompting techniques across all agents. that traditional application performance monitoring tools weren't designed to
handle.
Extend capabilities with Agent Skills. Agent Skills provide a structured way to
equip your agents with specialized knowledge, workflows, and tool integrations The core issue is that AI systems are non-deterministic with opaque reasoning
beyond their base capabilities. Rather than encoding all domain expertise processes. When an AI agent fails or behaves unexpectedly, you can't simply
directly in prompts, Skills act as modular capability packages that agents can examine a stack trace—you need visibility into prompt chains, model decision
leverage when needed. paths, retrieval contexts, token consumption, and the entire reasoning workflow.
Traditional debugging often falls short when the core logic happens inside a
Composable architecture: Skills can work together on complex tasks and
neural network.
invoke other skills as needed. This enables sophisticated workflows—for
example, a compliance skill might call a document analysis skill, which in turn The key insight is that debugging AI applications requires understanding not
uses a specialized extraction skill. This composability lets you build hierarchies just what happened, but why the model made specific decisions and how context
of capability without creating monolithic implementations. flowed through multi-step reasoning chains.

When to use Skills: With these foundational principles established, let's examine how they apply to
specific architectural patterns. We'll start with the simplest approach that works
• Domain-specific expertise (financial analysis, legal review, scientific research)
for the majority of enterprise use cases, then progress to more sophisticated
• Standardized workflows your organization has refined patterns that justify their complexity through measurable performance gains.

• Specialized tool integrations (databases, APIs, internal systems)

Single-agent systems
• Industry-specific best practices and compliance requirements
In a single-agent system, an AI-powered agent operates in a continuous loop:
Implementation approach: Skills integrate seamlessly with both single-agent perceive the environment, decide next steps, and act to accomplish a goal.
and multi-agent architectures. In single-agent systems, Skills extend the agent's
baseline capabilities. In multi-agent systems, different agents can configure Typically, interacting with an agent follows this pattern:
different skills based on their specialization—a financial analysis agent might
1. A user gives the agent a task.
use risk assessment skills while a customer support agent uses CRM integration
skills. 2. The agent then formulates a plan, executes actions based on available tools,
observes the results, and adjusts its approach based on feedback.
This modular approach means you can update skills independently without
rewriting agent logic, share Skills across multiple Agents, and scale capabilities 3. The agent keeps repeating this cycle until the task is completed or it hits a
as your organization's needs evolve. stopping condition like “pause here for human review."

11
 Example: Single-agent research agent

You deploy a research agent configured with the Model Context Protocol (MCP)
to connect the agent to your various systems, including content repositories,
business tools, and development environments. The diagram below shows how
your single agent can handle a relatively complex task leveraging multiple tools.

Architecture overview: The core components of a single-agent system include:

an AI model that serves as the reasoning engine, a prompt that defines the
agent's role and capabilities, and a toolkit of integrations that enable the agent to
interact with external systems and perform specific functions. Skills also provide
an additional layer of capability, equipping agents with specialized domain
knowledge, workflows, and best practices that extend beyond their base training,
enabling a single agent to handle complex tasks that might otherwise require
multiple specialized agents.

When to use: Single agents excel when tackling open-ended problems where
the path forward isn't clear from the start. You can't predetermine the solution 1. User query:
because you don't know how many steps will be required or what obstacles will An employee passes the following user query to the research agent:
emerge along the way. "Research remote work productivity tools that engineering teams are adopting

When to avoid: When you need to get the perfect answer on the first try, 100% of and see if any correlate with our internal productivity metrics"

the time. A single agent is a powerful resource, but to handle complex problems, 2. Initial analysis:
or to achieve the highest level of accuracy you will want to leverage multi-agent The Claude assesses the user query:
architecture. However, before scaling to multi-agent systems, consider whether • Think: "This query requires two distinct data sources: external research on
adding specialized skills to your single agent might achieve your accuracy productivity tools and internal company metrics. I should decompose this into
requirements more efficiently. parallel searches since the external research doesn't depend on internal data
initially. I'll need to correlate findings at the end."

12
3. Skills in action: This research agent leverages specialized skills, including: Database query successful—have baseline productivity metrics. Need refined
search for quantitative adoption data and team-specific feedback."
• Research methodology skills that guide systematic literature review
Claude executes follow-up queries based on this analysis:
approaches
• Data correlation skills that provide frameworks for identifying meaningful • Web search tool: Refined search targeting engineering-specific adoption
patterns patterns
• Business intelligence skills that ensure insights align with organizational
priorities • SQL database tool: Targeted queries correlating tool implementation periods
with productivity changes

these Skills enable the agent to apply proven frameworks rather than 7. Data synthesis and correlation: Using the think tool for comprehensive
reasoning from first principles, improving both accuracy and efficiency. analysis:

4. Task decomposition and planning: Think: "External research shows clear adoption trends across development

Based on the thinking analysis, Claude plans its approach: tools, project management platforms, and communication systems. Internal
metrics reveal productivity variations across teams and quarters. Cross-

• External web search for productivity tool adoption trends referencing implementation timelines with performance data to identify

• Internal database queries for company productivity metrics potential correlations while accounting for external factors."

• Parallel tool execution for optimal efficiency 8. Result generation: Claude synthesizes findings using extended context
• Correlation methodology for synthesis capabilities to maintain full conversation context and provides consolidated

5. Parallel tool execution: results:

Claude leverages its native parallel tool calling capability to execute multiple
tools simultaneously: “Research identified several categories of remote work productivity tools with
significant engineering team adoption: development environment tools…”

• Web search tool (via MCP: Searches for remote work productivity tools and
adoption data Multi-agent systems
• SQL database tool (via MCP): Queries internal productivity metrics across
teams and time periods Multi-agent architectures coordinate multiple specialized agents to tackle
complex problems that exceed the capabilities of a single generalist system.
Both tools execute concurrently, reducing total response time significantly. Rather than one AI model handling everything, tasks are decomposed,
distributed, and executed across multiple agents, often with distinct expertise for
6. Iterative analysis and refinement: After processing initial results, Claude specific types of queries. The results from multiple agents are then synthesized
uses the think tool for deeper analysis: into a coherent response.

• Think: "Web search returned comprehensive tool categories and general

adoption trends, but need more specific data on engineering team preferences.

13
Internal Anthropic research shows that for complex tasks requiring pursuit decisions multiply, it's essential to implement tracing that captures not just
of multiple independent directions simultaneously, multi-agent systems individual agent behavior but agent decision patterns and interaction structures
outperform single-agent systems by 90.2%. The key insight is that intelligence to diagnose root causes when coordination fails. Without comprehensive
reaches a threshold where "multi-agent systems become a vital way to scale observability into how agents communicate, delegate tasks, and synthesize
performance" because "groups of agents can accomplish far more" than results, debugging becomes nearly impossible when emergent behaviors arise
individuals, much like human organizations. from complex agent interactions.

Architecture overview: Multiple agents with specialized capabilities work When considering multi-agent implementation, start by clearly defining what
toward common goals. This might involve orchestrators delegating to you're trying to accomplish and build the simplest solution that meets your
specialists or hierarchical structures where senior agents manage subagents. requirements. Design for modularity and scalability from the beginning; you'll
Communication happens directly between agents or through shared memory appreciate this foundation when you need to add new capabilities or scale
and message queues that coordinate their efforts. Agent Skills can also be existing ones.
strategically distributed across agents to create deep specialization.
Architecture patterns
When to use: Multi-agent systems excel when single agents hit fundamental
limits. Choose multi-agent architectures when: (1) tasks involve open-ended Multi-agent systems organize around two fundamental coordination concepts:
problems where it's difficult to predict the required steps in advance and require centralized and decentralized architectures, each addressing different
the flexibility to pivot or explore tangential connections as the investigation coordination challenges and use cases.
unfolds; (2) you need specialized expertise that would overwhelm a generalist
agent, research shows single agents fall off sharply when there are two or Centralized systems employ hierarchical patterns where a central supervisor

more distractor domains; or (3) problems demand broad-based queries that intelligently delegates tasks to specialized agents, creating clear chains of

involve pursuing multiple independent directions simultaneously, where responsibility that mirror effective human organizational structures. These

parallel processing provides substantial performance gains. They're particularly hierarchical systems are known by various names such as supervisory,

effective for complex research, comprehensive analysis spanning multiple orchestrator, or router patterns, with each representing slightly different

disciplines, or scenarios requiring sustained autonomous operation across permutations of centralized control, while some focus primarily on task

diverse knowledge domains. delegation, others on routing decisions, and still others on full orchestration of
agent interactions.
Implementation considerations: Multi-agent systems deliver impressive
power for complex tasks, but that power comes with proportional complexity Decentralized systems use collaborative patterns where autonomous agents

in both architecture and operational costs. Multi-agent architectures consume communicate directly in peer-to-peer fashion, negotiate roles dynamically,

tokens rapidly, requiring tasks where the business value justifies the increased and solve complex problems through distributed intelligence. Collaborative

performance costs. Design your system to scale effort appropriately—simple systems are sometimes termed swarm or federated architectures, reflecting their

queries shouldn't trigger expensive multi-agent workflows. emphasis on emergent coordination rather than imposed control.

Observability becomes even more critical. As outlined previously, traditional Supporting these architecture patterns are agentic workflows, which provide

debugging approaches fail because agents make dynamic decisions and are structured orchestration for multi-step processes, defining the sequence and

non-deterministic between runs. In a multi-agent architecture where agent conditions under which agents execute tasks across distributed environments.

14
The core distinction lies in their coordination philosophy: centralized control Key challenge: Context management
versus distributed autonomy versus structured orchestration. Organizations
The orchestrator agent may face the fundamental problem that context grows
frequently combine these patterns to create robust, scalable solutions that match
too complex for one agent to manage effectively, creating performance
their specific business requirements.
bottlenecks as agents struggle to maintain coherence across extended
interactions. This context complexity manifests as context window overflow,
Hierarchical/supervisory systems
degraded reasoning performance, and coordination failures between agents.
Hierarchical systems use a central controller to coordinate multiple role-
Successful implementations need solid context management strategies: context
specific agents through intelligent task delegation. A supervisor agent analyzes
editing automatically clears stale tool calls and results when you approach token
incoming requests, routes them to appropriate specialists, and synthesizes
limits while keeping conversation flow intact, and memory tools let your agents
responses, creating a clear chain of responsibility that scales effectively with
store and retrieve information outside the context window through file-based
organizational complexity.
systems that persist across sessions. You should also consider having your
In hierarchical systems, individual subagents are treated as tools, where a tools include pagination, range selection, filtering, and truncation with sensible
supervisor agent uses a tool-calling model to decide which agent tools to invoke. defaults, capping responses at manageable sizes (something like 25,000 tokens)
This pattern mirrors how effective human teams operate: specialists focus to prevent context exhaustion.
on their domain expertise while coordinators handle task distribution and
integration. Subagents can also have their own subagents, with these groups
abstracted from the supervisor agent, which only interacts with the subagent
team leader and remains unaware of further delegation.

The economics favor this approach despite higher token usage. While multi-
agent systems consume significantly more tokens than single interactions, the
performance gains justify the cost for high-value, complex tasks that require
specialized knowledge or exceed single-agent context limits.

Implementation variations differ in how they balance coordination overhead

with response quality:

• Full orchestration systems maintain complete supervisory control over user

interactions and task execution

• Routing-focused implementations specialize in delegation decisions,

potentially handing off user communication to specialist agents

• Hybrid coordination approaches selectively involve supervisors based on

task complexity

15
Example: Multi-agent hierarchical workflow - Marketing campaign 2. Marketing director agent (supervisor): The supervisor agent analyzes
development the campaign requirements, identifies key deliverables, determines resource
allocation, and creates a strategic execution plan that maps specific tasks to
A marketing agency deploys a hierarchical multi-agent system to develop
appropriate specialist agents.
comprehensive marketing campaigns, with a supervisor agent coordinating
specialist agents to ensure strategic alignment while leveraging deep domain 3. Market research agent: Receives directive from supervisor to conduct target
expertise across all campaign components. audience analysis, competitive landscape research, and market opportunity
assessment, reporting findings back to the marketing director agent.

4. Creative design agent: Gets tasked by supervisor to develop visual concepts,

brand assets, and design frameworks based on market research insights
and brand guidelines, with the supervisor reviewing and approving creative
direction.

5. Copywriting agent: Receives assignment from supervisor to create messaging

strategy, ad copy, and content across all channels, ensuring consistency with
creative direction and market positioning established by previous agents.

6. Media planning agent: Gets directed by supervisor to develop media mix

recommendations, channel selection, budget allocation across platforms, and
timing strategy based on audience insights and creative requirements.

7. Campaign integration and approval: The marketing director agent is tasked

to synthesize all specialist outputs, ensures strategic coherence, resolves any
conflicts between specialist recommendations, and prepares the integrated
campaign proposal.

8. Submit comprehensive campaign strategy: The final integrated marketing

campaign with creative assets, media plan, budget allocation, timeline, and
success metrics is delivered to the client.

Collaborative systems
Collaborative systems enable multiple specialized agents to work together in
real-time through sophisticated coordination mechanisms, sharing information
1. Campaign brief submission: A client submits a marketing campaign brief and coordinating actions to achieve outcomes that exceed individual agent
including objectives, target audience, budget constraints, timeline, and brand capabilities. Unlike hierarchical systems with central control, collaborative
guidelines to the system. patterns emphasize peer-to-peer interaction where agents communicate

16
directly, negotiate roles dynamically, and collectively solve complex problems Example: Multi-agent collaborative workflow - Competitive Intelligence
through distributed intelligence. Gathering

In collaborative systems, agents operate as autonomous entities that

communicate, coordinate, and collaborate through various mechanisms to
achieve collective goals. This approach mirrors human teamwork where
specialists contribute their expertise while maintaining awareness of the
broader objective. The key differentiator is that coordination emerges from agent
interactions rather than being imposed by a central authority.

Implementation variations differ in how agents coordinate and share context:

• Group chat orchestration enables multiple agents to solve problems, make

decisions, or validate work by participating in a shared conversation thread
where they collaborate through discussion

• Event-driven coordination uses events as a shared language, acting as

structured updates that enable agents to interpret instructions, share context,
and coordinate tasks

• Blackboard architectures provide shared knowledge repositories where all

agents can read from and write to a central knowledge repository acting as
collective memory

Key challenge: Communication complexity and emergent behavior

unpredictability

Collaborative systems face fundamental challenges in managing inter-agent

communication and predicting system behavior. Frequent communication
between agents leads to increased computational costs and complexity, while
multi-agent systems have emergent behaviors that arise without specific
programming, where small changes can unpredictably affect how agents
behave. Success requires frameworks for collaboration that define division
of labor, problem-solving approaches, and effort budgets rather than strict
instructions. Additional challenges include preventing agents from bouncing
tasks indefinitely and implementing robust conflict resolution mechanisms.

17
A strategic consulting firm deploys a collaborative multi-agent intelligence Agentic workflows
system where specialized analysis agents work together in real-time, cross-
referencing findings and building comprehensive competitive landscapes that Agentic workflows define the structure of how agents operate, including how
exceed individual agent capabilities through collective intelligence. they communicate, hand off tasks, and collaborate toward shared objectives.
Unlike the dynamic behavior of individual agents, workflows are predefined
1. Intelligence request initiation: Client requests comprehensive competitive and static. The two common agent workflow patterns are sequential and
analysis, triggering coordinated intelligence gathering across all specialized hierarchical.
analysis agents.

2. Coordinated data collection: The client request is placed in a queue. Pricing, Sequential workflows
product, marketing, financial, social media, and strategic intelligence agents
Sequential workflows use predetermined control flow with defined execution
establish communication channels and divide monitoring responsibilities to
paths, ensuring predictable agent transitions that are ideal for repeatable
avoid duplication.
processes like document approval chains or compliance checks. These workflows
3. Cross-agent collaboration: Agents continuously share findings in real-time - provide clear audit trails and deterministic behavior, making them well-suited
pricing agents alert product agents about feature-price correlations, marketing for regulatory environments where process consistency and traceability are
agents share campaign data with financial agents, and social media agents critical.
provide sentiment insights to strategic agents.
Sequential workflows can leverage either software-defined decision points, such
4. Intelligence validation: All agents cross-reference discoveries to identify as conditional logic based on task outcomes or system state changes, or AI-driven
contradictions, validate findings across multiple data sources, and build routing where models decide application control flow based on intermediate
corroborated competitor profiles. results and contextual factors. This hybrid approach allows for both the reliability
of predetermined paths and the flexibility to adapt based on content analysis or
5. Collective synthesis: Agents collaborate to integrate multi-dimensional dynamic conditions.
insights, assess market opportunities, and develop predictive intelligence about
competitor strategic moves. The key advantage is operational predictability, you can map out the entire
process flow, estimate execution costs, and debug issues by examining specific
6. Strategic intelligence report: The report agent develops a comprehensive workflow stages. However, this predictability comes at the cost of flexibility
competitive landscape analysis with validated findings, predictive insights, and when handling edge cases or novel scenarios that don't fit the predefined
strategic recommendations based on collective agent intelligence. workflow structure.

7. Submit intelligence report: The final integrated intelligence report with When to use: Use sequential workflows when tasks can be cleanly decomposed
metrics is delivered to the client. into fixed subtasks. The main goal is to trade off latency for higher accuracy by
making each AI call an easier, more focused task.

Consider sequential orchestration in these scenarios: multi-stage processes

with clear linear dependencies and predictable workflow progression, data
transformation pipelines where each stage adds specific value that the next
stage depends on, workflow stages that can't be parallelized, and progressive

18
refinement requirements like draft-review-polish workflows. A company deploys a multi-agent workflow solution to automate their data
analysis requests, enabling rapid insights generation without bottlenecking the
Use sequential patterns when you understand the availability and performance
data science team.
characteristics of every agent in the pipeline, and where failures or delays in one
agent's processing are tolerable for the overall task completion. 1. Analysis request: A stakeholder submits a data analysis request through the
system (e.g., "Analyze Q4 sales performance by region" or "Identify customer
Examples where sequential workflows prove effective include generating
churn risk factors").
marketing copy then translating it into different languages, writing a document
outline, validating that outline meets specific criteria, then writing the full 2. Scoping agent: The scoping agent analyzes the incoming request, determines
document based on the approved outline. the analysis type (descriptive, diagnostic, predictive, or prescriptive), identifies
required data sources and methodologies, assesses complexity level, and routes
When to avoid: Avoid sequential workflows for processes that include only
to the appropriate analytical pathway.
a few stages that a single agent can accomplish effectively, when agents
need to collaborate rather than hand off work, or when the workflow requires 3. Data engineering agent: The data engineering agent uses the scoping output
backtracking or iteration. to extract data from relevant sources (data warehouses, APIs, databases),
performs data cleaning and validation, handles missing values and outliers,
Example: Multi-agent sequential workflow - automated data engineers relevant features, and prepares analysis-ready datasets.
science insights
4. Analysis agent: The analysis agent takes the prepared data and executes the
appropriate analytical workflow—running statistical tests, building models,
generating visualizations, identifying key patterns and insights—or flags
complex requests requiring human data scientist intervention with a detailed
handoff package.

5. Review/escalation: The analysis results are either automatically validated

and approved for distribution or queued for review by a senior data scientist for
quality assurance and interpretation refinement.

6. Deliver insights: The final analysis outputs (reports, dashboards, model

predictions, or recommendations) are packaged and delivered to stakeholders
through their preferred channels.

19
Parallel workflows other's work or require cumulative context in a specific sequence, when the task
requires a specific order of operations or deterministic results from running
Parallel workflows distribute independent tasks across multiple agents in a defined sequence, or when resource constraints like model quotas make
simultaneously, with results merged or processed concurrently. This pattern parallel processing inefficient. Don't use parallel patterns when agents can't
excels when tasks require diverse perspectives or specializations, enabling reliably coordinate changes to shared state or external systems while running
significant speed improvements through concurrent processing. simultaneously, when there's no clear conflict resolution strategy to handle
The concurrent orchestration pattern runs multiple agents simultaneously on the contradictory results from each agent, or when result aggregation logic is too
same task, allowing each agent to provide independent analysis from its unique complex or lowers the quality of the results.
perspective or specialization. This resembles the fan-out/fan-in cloud design
pattern, where results are often aggregated but not required to be. Example: Multi-agent parallel workflow - financial risk
assessment
This pattern addresses scenarios requiring diverse insights or approaches to
the same problem. Instead of sequential processing, all agents work in parallel, A financial institution deploys a multi-agent parallel workflow to evaluate loan
reducing overall runtime and providing comprehensive problem space coverage. applications and investment opportunities, enabling faster decision-making
Each agent can independently produce results within the workload, such as while maintaining comprehensive risk analysis across critical dimensions.
invoking tools or updating different data stores.

Agents operate independently without handing off results to each other, though
an agent might invoke additional agents using its own orchestration approach.
The pattern supports both deterministic calls to all registered agents and
dynamic selection based on task requirements.

When to use: Use parallelization when divided subtasks can be processed

simultaneously for speed, or when multiple perspectives are needed for higher
confidence results. For complex tasks with multiple considerations, AI models
generally perform better when each consideration is handled by a separate call,
allowing focused attention on each specific aspect.

Examples where parallelization proves useful include sectioning approaches like

implementing guardrails where one model processes user queries while another
screens for inappropriate content, or automating evaluations where each call
evaluates different aspects of model performance. Voting patterns work well
for reviewing code vulnerabilities with several different prompts, or evaluating
content appropriateness with multiple prompts using different vote thresholds
to balance false positives and negatives.

When to avoid: Avoid parallel workflows when agents need to build on each

20
1. Risk assessment request: A loan application or investment proposal is Evaluator-optimizer
submitted to the system for comprehensive risk evaluation.
Evaluator-optimizer workflows use two AI systems in iterative cycles, one
2. Data aggregation agent: The data aggregation agent collects all relevant generates content while another evaluates and provides feedback, repeating
information including credit reports, financial statements, market data, until quality standards are met. This pattern delivers significant improvements
regulatory filings, and historical performance metrics from internal and external when properly implemented, though it comes with higher token costs.
data sources.
The pattern operates through structured feedback loops where a generator
3. Parallel agents (leveraging their own tools) creates initial responses and incorporates feedback for successive improvements,
while an evaluator assesses content against predefined criteria and provides
3a. Credit risk agent: Simultaneously analyzes borrower
actionable guidance. This resembles writer-editor collaboration, with specific
creditworthiness, debt-to-income ratios, payment history, and collateral
suggestions incorporated in revised drafts.
quality to generate credit risk scores and probability of default
calculations. When to use: Use evaluator-optimizer workflows when clear evaluation
criteria exist and iterative refinement provides demonstrable value through AI
3b. Market risk agent: Concurrently evaluates market volatility, interest
feedback loops. This pattern excels for content creation requiring nuance like
rate sensitivity, sector exposure, and economic indicators to assess
literary translation, code generation with security requirements, professional
potential losses from market movements and economic downturns.
communications where tone matters, and research tasks needing multi-step
3c. Operational risk agent: In parallel, examines internal process reasoning with validation.
risks, fraud indicators, compliance gaps, and operational capacity to
When to avoid: Avoid evaluator-optimizer workflows when first-attempt quality
handle the transaction, identifying potential operational failures or
already meets requirements, evaluation criteria are subjective or unclear, or
irregularities.
when time and cost constraints outweigh quality improvements. Don't use
3d. Regulatory compliance agent: Simultaneously reviews regulatory this pattern for real-time applications requiring immediate responses, simple
requirements, anti-money laundering checks, know-your-customer routine tasks like basic classification, or resource-constrained environments with
compliance, and jurisdictional restrictions to ensure full regulatory strict token budgets. Avoid when deterministic solutions exist, when evaluator
adherence. workflows lack domain expertise for meaningful feedback, or when performance
degradation outweigh benefits.
4. Risk aggregation and decision engine: All parallel risk assessments are
consolidated, weighted according to institutional policies, and synthesized into
comprehensive risk profiles with actionable recommendations.

5. Submit risk assessment results: The final multi-agent risk evaluation with
approval/denial recommendations, risk scores, and detailed analysis reports is
delivered to decision makers.

21
Example: Multi-agent evaluator workflow - API documentation 4. Refinement cycle: Generator incorporates feedback from both evaluators and
creator iteratively improves documentation until all criteria are met.

A software development organization deploys an evaluator-optimizer workflow 5. Published documentation: Final polished API documentation is
to automatically generate comprehensive API documentation from codebases, automatically published to the developer portal with interactive examples and
ensuring technical accuracy and developer usability through iterative comprehensive reference materials.
refinement cycles that eliminate manual documentation bottlenecks.
This process typically runs 2-4 cycles, significantly improving documentation
quality while maintaining technical accuracy.

Emerging patterns
As organizations push the boundaries of what's possible with AI agents,
several experimental patterns are moving from research labs into early-stage
implementations. In the following section, we highlight some of these emerging
patterns.

Agent pattern: Dynamic agent generation

Dynamic agent generation represents a burgeoning experimental approach

that takes modularity to its logical conclusion: agents created at runtime by
assembling components from libraries of prompts, tools, and configurations,
then dissolved after task completion. While no production systems currently
implement true dynamic creation, the technical foundations exist across multiple
research projects and experimental frameworks like AutoGen or Semantic
Kernel.
1. Code input: Development team submits API codebase to the documentation
generation system. This pattern offers compelling advantages for resource optimization and
task-specific performance; rather than maintaining pre-configured agents,
2. Generator agent: Analyzes codebase and creates initial documentation
systems can analyze incoming requests and automatically instantiate agents
including endpoint descriptions, parameters, examples, and authentication
with precisely the required capabilities, then free those resources when tasks
requirements.
complete. However, significant challenges remain around context management
3. Technical evaluator agent: Validates documentation accuracy against actual complexity, emergent behavior risks, and the overhead costs of dynamic
code implementation, checking parameter types, endpoint coverage, and creation. Current research into multi-agent coordination and event-driven
example correctness. architectures provides the groundwork, but organizations should approach this
as experimental territory.

22
Architecture pattern: Network/peer-to-peer systems For scenarios where you need flexibility but still want oversight, hierarchical
systems give you the best of both worlds. A supervisor agent can enforce
Network architectures represent significant evolution in multi-agent business rules while specialist agents handle the complexity.
coordination, eliminating hierarchical bottlenecks through "many-to-many
agent communication where any agent can communicate with any other agent Low control requirements (research, brainstorming, complex analysis) →
directly". Early benchmarking shows "swarm architecture slightly outperforms Collaborative multi-agent systems become viable
supervisor architecture across the board" because agents can collaborate directly
When the goal is exploring possibilities or handling truly complex problems, the
without supervisory translation layers.
unpredictability of collaborative agents becomes a feature, not a bug.

2. How complex is your problem domain?

Decision framework:
Which pattern for which use case Single domain problems (answering product questions, processing returns,
generating reports) → Single agents handle these efficiently
Understanding architecture patterns is only the first step. The critical challenge
for engineering leaders is selecting the right approach for your specific Don't over-engineer. If the work involves straightforward, repeatable tasks, a
constraints: budget, timeline, complexity, and risk tolerance. Rather than single well-designed agent is likely sufficient.
choosing based on technical sophistication, successful implementations match
Multi-domain but predictable problems (employee onboarding, compliance
architectural complexity to business value through systematic evaluation of
workflows, standard analysis tasks) → Sequential or parallel workflows
three key dimensions.
When you can map out the process steps but need different expertise at each
The three critical questions stage, workflows provide structure without excessive complexity.

Before diving into specific patterns, every enterprise team needs to answer these Complex, open-ended problems (strategic analysis, research projects, system
fundamental questions: troubleshooting) → Multi-agent architectures

1. What level of control do you need? These problems require breaking down into smaller parts and different
approaches. If the work benefits from multiple perspectives or specialized Skills,
High control requirements (regulatory compliance, financial transactions,
multi-agent systems may make sense.
safety-critical operations) → Start with single agents or sequential workflows
3. What are your resource constraints?
Think about it this way: if you need to explain exactly why the system made a
specific decision to auditors, regulators, or executives, you want predictable, Limited budget/tokens → Single agents or carefully designed parallel
traceable behavior. A single agent handling loan approvals with clear decision workflows
criteria is far easier to audit than a multi-agent system where three different AI
Multi-agent systems use roughly 10-15x more tokens than single agents. Do the
models collaborated on the recommendation.
math on your expected volume before committing to complex architectures.
Moderate control requirements (customer support, content creation, data
Time-to-market pressure → Start with single agents, plan an evolution path
analysis) → Consider hierarchical multi-agent systems

23
You can deploy a single agent in weeks. Multi-agent systems take months to get • Code review and basic development tasks
right. Build something that works, then enhance it.
• Routine analysis and reporting
Long-term strategic initiative → Design for modular evolution
Sequential workflows work best for:
If this is a multi-year initiative, build your first single agent with interfaces that
• Multi-step approval processes
support adding more agents later. Design for evolution from the beginning:
maintain consistent user experiences while building capability for backend • Content creation pipelines (draft → review → publish)
architectural changes as requirements grow.
• Data transformation and validation
4. Do you need deep domain expertise?
• Compliance checking with multiple criteria
Single domain with established workflows → Single agent with specialized
Skills Parallel workflows work best for:

Before jumping to multi-agent architectures, consider whether a single agent • Multiple perspectives improve quality

equipped with domain-specific skills can solve your problem. Skills provide deep • Independent analyses can run simultaneously
expertise without the complexity of multi-agent coordination.
• Speed matters more than coordination overhead
Multiple distinct domains requiring coordination → Multi-agent systems
with specialized Skills • Risk assessment requires diverse viewpoints

When domains must work together (e.g., legal review coordinating with financial Multi-Agent systems work best for:
analysis), multi-agent systems where each agent has appropriate Skills provide
• Complex problem-solving requiring diverse expertise
both specialization and coordination.
• Research and analysis projects
Example: A contract review system might start with a single agent using legal
skills. As complexity grows, it could evolve into a multi-agent system where • Dynamic customer interactions spanning multiple systems
separate agents handle contract analysis, risk assessment, and compliance
• Strategic planning and decision support
checking—each with their own specialized Skills.

Pattern selection guide

Real-world evolution: An e-commerce platform's journey
These constraints translate into clear architectural recommendations:
• Phase 1: Single agent for customer inquiries (proving value)
Single agents work best for:
• Phase 2: Routing pattern separating order status, product questions,
• Customer service processes for well-defined product categories complaints

• Document processing with clear business rules • Phase 3: Specialized agents for each category with shared context

24
• Phase 4: Multi-agent system with inventory, payment, and shipping Remember, you are not bound by the simple architecture patterns. When your
coordination business needs justify the added complexity, combining patterns strategically
can unlock capabilities that single approaches cannot achieve.
• Phase 5: Evaluator agents for quality assurance and continuous improvement

The key: your architecture should evolve with your needs. Start simple,
measure everything, add complexity only when it delivers measurable value.
The best architecture is the simplest one meeting today's requirements while
providing a path to tomorrow's capabilities.

Hybrid architecture strategies

The decision framework provides clear starting points, but production

systems often evolve into hybrid architectures that combine multiple patterns
strategically. Understanding these combinations prevents architectural dead
ends and enables systematic scaling.

Common hybrid patterns:

Hierarchical systems with parallel processing

• A supervisor agent delegates to specialist agents; the specialist agents

coordinate parallel workflows. For example, a financial risk assessment
supervisor might delegate to credit, market, and operational risk agents, each
running parallel analyses within their domain.

Sequential workflows with dynamic routing

• Linear processes that invoke different agent types based on intermediate

results. A customer service workflow might start with classification, then route
to either a simple resolution agent or a complex multi-agent research team
based on issue complexity.

Single agents with multi-agent escalation

• Simple agents handle routine tasks but automatically trigger sophisticated

multi-agent systems when encountering edge cases. This optimizes costs
while maintaining capability for complex scenarios.

25
Chapter 4

Looking forward: the future

of building AI agents 26
 In this guide we’ve examined the full spectrum of AI agent implementations,
Chapter 4
from single-agent systems handling focused tasks to sophisticated multi-agent
architectures tackling complex, multi-domain challenges. We’ve seen real-world

Looking forward: use cases across industries, explored architectural patterns and their trade-
offs, and finally established frameworks for making informed implementation

the future of building

decisions. With this foundation you can confidently build AI agents that solve
real problems and deliver real results.

AI agents
Successfully implementing AI agents requires aligning technical complexity
with business value rather than chasing the most sophisticated architecture
you can build. You'll see the best results if you start with single agents to prove
ROI, build observable systems from day one, and evolve your architecture based
on what the data tells you. Organizations that take this measured approach
consistently outperform those that over-engineer right out of the gate. The
frameworks and patterns we've covered give you a solid foundation, but your
specific implementation will depend on your risk tolerance, resource constraints,
and how ready your organization is for autonomous systems.

The organization that can rapidly iterate between simple and complex
approaches as business requirements evolve is the organization that can win.
Whether you deploy a single customer service agent or orchestrate multi-agent
research systems, your North Star must be modular designs, comprehensive
observability, and clear success metrics that connect directly to business
outcomes.

The tools are ready, the playbook is written. Now it's time to solve real-world
problems.

27
Chapter 5

Next steps 28
 Chapter 5

Next steps
Ready to build? Get started with the Claude Developer Platform to access
the models, tools, and technical documentation you need for AI agent
implementation. Whether you're prototyping your first single-agent system
or scaling to multi-agent architectures, these resources will accelerate your
development:

• Start building agents with the Claude Developer Platform - Access

comprehensive API documentation, implementation guides, and prompt
engineering techniques.

• Explore Agent Skills - Learn how to equip your agents with specialized
knowledge, workflows, and tool integrations.

• Watch Building the future of agents with Claude - Deep dive into advanced
architectures and emerging pattern with the leaders who built the Claude
Developer Platform.

• Explore the Anthropic Engineering Blog - Technical articles on agent

development, context engineering, and production deployment strategies.

Contact our Sales team to learn more or sign up with the Claude Developer
Platform today.

29
[Link]