04Open-source agent framework
Overview
Open-source AI agent framework with human-in-the-loop oversight. Agents follow a structured Plan → Act → Reflect → Evaluate loop with approval gates that pause execution before sensitive operations. Features multi-agent teams (route, broadcast, delegate modes), custom tool registry with Pydantic validation, extensible Letta-style memory, sandboxed execution, and lightweight LLM wrappers replacing LangChain with direct Anthropic/OpenAI SDK integration. Configurable via YAML, runs as CLI or FastAPI server.
The challenge
Agent systems can become difficult to inspect once planning, tool use, memory, and delegation are hidden behind abstractions. Sensitive actions also need a deliberate human checkpoint.
The approach
Pori makes the control loop explicit. Agents plan, act, reflect, and evaluate while typed tools, approval gates, and routing modes keep execution understandable and interruptible.
01Process
How it works
- 01
Plan
Turn the goal into an explicit sequence of actions.
- 02
Act
Run validated tools through the configured registry.
- 03
Reflect
Inspect intermediate work and update the working context.
- 04
Evaluate
Check the result, route follow-up work, or request approval.
02Decisions
Key decisions
- 01
Keep orchestration visible
The agent loop is represented directly instead of being buried inside a large orchestration dependency.
- 02
Pause before sensitive work
Human-in-the-loop gates stop execution at meaningful boundaries so a person can inspect and approve the next action.
- 03
Use direct model adapters
Lightweight OpenAI and Anthropic wrappers replace LangChain while preserving a small, understandable integration surface.
Technologies
- Python
- FastAPI
- Anthropic
- OpenAI
- Pydantic
- Multi-Agent
- HITL