DoStuff

AI Agent Harness

• reAct tool calling
• short term memory
• long term memory
  • Procedural
  • Semantic
  • Episodic
• context management
• agent skills
• mcp servers (stdio and http support)
• self-learning loop
• litellm python sdk
• opentelemetry tracing

Setup

• run git clone https://github.com/kVarunkk/DoStuff.git
• create a .env in the root and refer .env.example for the variables
  • Refer to litellm supported models and the official docs for more info.
• create and activate a venv
• run pip install -r requirements.txt
• create a DOSTUFF.md for giving your agent a personality. Take inspiration from DOSTUFF.sample.md
• run python app.py

Tools

• write_file
• read_file
• list_files
• delete_file
• run_code

Guidelines

• Main agent code is present in agent/run_agent.py
• To create a new tool:
  • add a new file in the tools directory with name matching that of the tool func
  • update tools/definitions.py
• New files will be created in the agent_workspace directory
• Add skills in the skill directory. Sample skills present. Refer this for more info.

Short Term Memory

• Uses SQLite for Storage, In-Memory Storage also available.

Long Term Memory

• Procedural: Saved as skills. Learning loop creates/updates skills at the end of the session.
• Semantic: Saved in Vector DB. User scoped. (To test, use the same user id)
• Episodic: Saved in Vector DB. User and Session scoped. (To test, use the same user id and session id)

MCP Servers

• Create and add mcp servers in mcp_config.json. Sample remote and local servers are present in mcp_config_sample.json.

Self-learning Loop

• At the end of each session on exit, a seperate agent loop decides if there is something in the conversation history worth learning from. If present, the agent creates a new skill for that and updates if the skill is already present using the skill-creator skill present in the /skills directory.

Observability

• Use this command to spin up a Jaeger container: docker run -d --name jaeger ` -e SPAN_STORAGE_TYPE=badger ` -e BADGER_EPHEMERAL=false ` -e BADGER_DIRECTORY_VALUE=/badger/data ` -e BADGER_DIRECTORY_KEY=/badger/key ` -v ${PWD}/jaeger_data:/badger ` -p 16686:16686 ` -p 4317:4317 ` -p 4318:4318 ` jaegertracing/all-in-one:latest
• Open http://localhost:16686 on your browser to view live traces.