Here's a project tree that maps one-to-one onto the agent roster and state contract from before. I'm assuming a Python deepagents backend (the primary package) served through LangGraph server, with a separate JS/Next frontend that talks to it over the LangGraph SDK — that split gives you streaming, thread persistence, and the human- approval interrupts essentially for free. slidegen/ |— README. md [_ backend/ (Le ITTTTTTTTI | || [= | [= 1 | = I | = | docker-compose. ymt |— .[Link] - [Link] | F [Link] / Dockerfile [L sre/slidegen/ }_ _init_.py [+ [Link] [- config. py | [Link] | | K orchestrator _init_.py prompt. py | subagents/ —init_.py ‘template_analyst.py document_modeler. py eae # # out Line_architect.py slide_mapper. py content_tailor. py asset_generator. py slide_renderer. py # backend + frontend; backend image carries Li deepagents, langgraph, langchain, fastapi, python-pptx, markitdown[pptx], pillow, pyd registers the deepagent graph for ~langgraph apt: libreoffice, poppler-utils (soffice + build_agent() -> create_deep_agent(...) ; th model names, MAX_QA LOOPS, render DPI, capac state extensions: run_id, workspace_path, st phase sequencing + when to terminate the QA each module = {name, description, prompt, to SUBAGENTS = [template_analyst, document_mode invoked in parallel (async subagents), one p visual_qa_inspector.py # spawned fresh each QA pass (fresh eyes) [Link] [- tools/ —init_.py schemas. py catalog_builder. py ‘thumbnail. py pptx_clone_edit. py pptx_pack. py render_slides. py extract_text. py assets. py He RRS HHH deterministic; LLMs CALL these, never emit r pydantic: Catalog, ContentModel, Outline, Ma SlideContent, QAReport (the filesystem co unpack XML -> placeholder bbox / fonts / cha 9-up grid for the catalog overview clone template slide + placeholder replace + pack/unpack wrappers soffice -> pdf -> pdftoppm 1-up @150dpi (per markitdown + grep for xxxx|Lorem| ipsum image search/gen, icons, chart builder from[LL skitts/ 1 | |‘ pptxe | backends/ L_ filesystem. py L apis I | HL _init_. | [Link] 1 | L uploads. py I | | workspace/ L {run_idby inputs/ catalog. json content_model.json out Line. json mapping. json py assets/ TTTTTTTTI ga_report. json L_ deck. pptx L tests/ [/ test_catalog_builder.py | test_clone_edit. py | L_ fixtures/ frontend/ | package. json [LL .env. local. example L- apps [— layout. tsx L_ page. tsx | components/ | Chat. tsx | [Link] |_ PlanTimeLine. tsx | Approvatate. tsx | SlidePreview. tsx ‘_ DeckDownload. tsx _ vbs | [Link] L uploads. ts slides/slide_NN. json renders/slide-NN. jpg # ea vendored Anthropic pptx skill ([Link] + sc deepagents FS backend rooted at workspace/{r FastAPI: upload, kickoff run, download deck writes template. pptx + docs -> workspace/{ru gitignored; the agents' virtual filesysten, template. pptx, source documents the riskiest tool — test it hardest next, @langchain/langgraph-sdk, react NEXT_PUBLIC_LANGGRAPH_URL, NEXT_PUBLIC_UPLOA useStream(): tokens + tool calls + interrupt «pptx template + documents dropzone -> uploa renders the write_todos plan as it updates resumes interrupts at outline / mapping chec shows renders/slide-NN. jpg as QA produces th final deck. pptx A few wiring decisions worth calling out, since they're where this kind of project usually goes sideways: graph. py is the whole system. It calls create_deep_agent with the orchestrator prompt, the shared tools/ , the SUBAGENTS list, the pptx skill, and the filesystembackend. langgraph. json points at it, so langgraph dev serves it with streaming and persistence, and the frontend's useStrean consumes it directly — no custom websocket layer. The run_id binds the chat thread to a workspace. When the user uploads via the FastAP! endpoint, you mint a run_id , drop files into workspace/{run_id}/inputs/ , and root the deepagents filesystem backend at workspace/{run_id} . Every subagent then reads/writes plain relative paths ( catalog. json , slides/slide_03. json ) and the orchestrator never has to carry slide content in its own context — that's the whole point of the filesystem backend. [Link] is the contract, not decoration. Each subagent's only job is to read its input files and emit one validated JSON artifact. Validate on write so a malformed mapping. json fails loudly at phase 3 instead of silently corrupting phase 5. System deps are a real gotcha. soffice and pdftoppm are not pip-installable — they must be in the backend Docker image. If you run agents in deepagents' sandbox backend instead of the host, the sandbox image needs them too. Human checkpoints ride on LangGraph interrupts. The outline and mapping gates are interrupt() calls inthe orchestrator; [Link] catches them from the stream and resumes with the user's edit. That's why the frontend talks to LangGraph directly rather than a plain REST wrapper — you'd lose interrupts otherwise. Want me to scaffold the actual files — [Link] , one representative subagent (say slide_renderer.py ), and [Link] — so you have a runnable skeleton to build out?