Stateful Agent Runtime

Planning, memory, tools, and review in one inspectable loop.

Agent Kernel Lab is a small Python project that shows the core plumbing of an Agent system without hiding it inside a prompt.
中文：这是一个小型 Python Agent runtime，重点展示规划、记忆、工具调用、决策路由和多 Agent review 的底层结构。

Run locally View trace

Architecture

Separate modules for the parts that usually become implicit.

中文：把 Agent 系统里容易混在一起的部分拆成可测试模块。

TaskDecomposerturns broad goals into task objects

DecisionEngineroutes task kinds to agent roles and tools

MemoryStorestores short and long-term context with ranked retrieval

ToolRegistryexecutes explicit local tools and handles unknown tools

StateMachinekeeps runtime transitions valid and auditable

AgentExecutorwrites outputs back into memory and trace

Execution Trace

A run produces decisions, tool calls, memory hits, and state transitions.

The project is deterministic by default, which makes it easier to test and inspect.

Goal: Design an AI Agent system...
Final: Completed 6 tasks using builder, critic, planner

Tasks
- task-1: Clarify objective and constraints
  agent=planner tool=extract_keywords
- task-2: Handle requirement slice 1
  agent=builder tool=implementation_plan
- task-6: Review plan risks and missing context
  agent=critic tool=risk_review

Run

Install, execute, and verify from the terminal.

git clone https://github.com/jsdnaasd/agent-kernel-lab.git
cd agent-kernel-lab
python3 -m venv .venv
source .venv/bin/activate
pip install -e ".[dev]"

agent-kernel run "Design an AI Agent system with planning, memory, tool calling, and multi-agent review"
pytest -v