Instruction file imported from tituslhy/fictional-bassoon (.cursor/rules/deep-agent-scope-lock.mdc). Copyright stays with the author.

Deep Agent Scope Lock

Context

backend/src/agent.py builds the agent via create_deep_agent (from the deepagents package) with a single tool — TavilySearch(max_results=5) — and init_chat_model(model="openai:gpt-5.4-nano"). That minimalism is intentional: the engineering investment in this project is the distributed pipeline around the agent, not the agent's own reasoning surface.

One correction against CLAUDE.md: it currently claims agent.py "constructs the LangGraph agent at module level — no factory wrappers." The actual code is two functions, create_agent() and get_agent() — not a module-level constant. Whether that's stale documentation or an earlier refactor that didn't get written back, treat the code as ground truth, not the doc, and fix the doc to match once this rule is settled.

Hard rules

• This rewrite changes how the agent's output is transported (AG-UI), rendered (A2UI), and exposed externally (A2A). It does not change what the agent does.
• Do not add tools, expand past the single Tavily tool, restructure the graph, or touch model= / temperature= as a side effect of protocol wiring.
• If an AG-UI or A2A requirement seems to force a change to create_agent() or get_agent() itself (not just how their output is consumed), stop and flag it to Titus rather than resolving it inline.