tinyagent

A generalised planning agent. Basically taking what I found most incredible about Claude Code and rebuilding it. This is meant as a hobbie project. Includes:

• generalised planning and execution architecture, task breakdown, tool usage
• terminal UI (react, ink)
• local model support (ollama)

Status.

Works for now.

How does it work?

The architecture is split into:

• Planner. Creates a plan from a goal.
• Executor. Executes a single step, runs tools, returns outputs, accumulates context.
• Completion validator. Validates whether a step is complete or not, based on executor's context.
• Orchestrator / state machine. Manages the whole flow.

It uses under the hood:

• qwen2.5-coder:14b
• JSON to parse the plans
• Ollama as an LLM inference server.
• ReAct-style XML with , , and tags to implement tool calling.

References:

• https://arxiv.org/pdf/2210.03629