agent-profiles

name	agent-profiles
description	Build or modify LangChain, DeepAgent, DeerFlow profiles, agent tools, middleware, checkpointing, skills wiring, and the shared harness layer in genai-tk.

GenAI Toolkit Agent Profiles

Read First

• docs/agents.md
• docs/deer-flow.md
• docs/middleware-pii-and-routing.md
• genai_tk/agents/langchain/config.py
• genai_tk/agents/langchain/factory.py
• config/agents/langchain/defaults.yaml

Configuration Shape

LangChain profiles are dict-keyed. The key is what users pass to -p.

langchain_agents:
  research:
    name: Research
    type: deep
    llm: gpt_41@openai
    tools:
      - spec: web_search
    skills:
      directories:
        - ${paths.project}/skills

Code Map

Concern	Paths
LangChain profiles and factory	genai_tk/agents/langchain/
Middleware	genai_tk/agents/langchain/middleware/
Tool specs and factories	genai_tk/agents/tools/
SmolAgents CLI	genai_tk/agents/smolagents/
DeerFlow bridge	genai_tk/agents/deer_flow/
DeepAgent CLI bridge	genai_tk/agents/deepagent_cli/
Sandbox backend	genai_tk/agents/sandbox/, genai_tk/agents/langchain/sandbox_backend.py

Change Workflow

1. Decide whether the change is profile-only, a new tool factory, middleware, or agent runtime behavior.
2. For profile-only changes, edit config/agents/**.yaml and verify with uv run cli agents langchain --list.
3. For tools, expose a factory under genai_tk/agents/tools/... and reference it from YAML.
4. For middleware, add a Pydantic config model if the YAML accepts options.
5. Add structural tests under tests/unit_tests/agents/ and integration tests only for real agent behavior.

Commands

uv run cli agents langchain --list
uv run cli agents langchain -p simple "What is 2+2?"
GENAITK_PROFILE=pytest uv run pytest tests/unit_tests/agents -q

Avoid

• Do not use display name where the profile key is required.
• Do not hardcode system prompts in Python when they belong in YAML.
• Do not pass credentials through normal tools; use secure credential tooling for browser profiles.