Build Your Own Agent

Problem

Many engineers can use Agent frameworks but cannot explain why a ReAct loop works, how tool results are folded back into context, or why multi-agent systems can contaminate each other. build-your-own-agent is aimed at the “understand Agents from first principles” use case: start with a minimal ~80-line Agent, then extend tools, memory, planning, resilience, and collaboration.

How It Works

The skill follows a ten-stage curriculum and loads references/0X-*.md only when needed:
- Stages 0-1: build the Agent-vs-LLM mental model and a first ReAct loop.
- Stages 2-4: add tool registration, three-layer memory, and planning with Schema generation, dependency analysis, and replanning.
- Stages 5-7: expand to a multi-tool research assistant, retries/timeouts/degradation, and predictable state-machine workflows.
- Stages 8-10: cover multi-agent collaboration, evaluation, and experience-based self-improvement.

It supports free exploration, guided mentorship, challenge mode, and speed-run mode, and tracks progress in progress.md. By default it operates in review mode: high-risk or irreversible actions stop for confirmation, and failed external calls retry up to three times with readable errors.

Boundaries

This is a teaching skill, not a rapid project generator. If the goal is fast output, finish the core stages first and then use framework-based tools. It also does not make consequential decisions, such as wiring in production credentials or executing non-reversible external actions.