Build the loop, the tools and the guardrails — and make your own coding agent that reads, plans, edits, tests and repairs its way to a verified pull request. 100% offline, zero API keys.
What we covered
The model is only half the agent. We built the other half — the harness around the LLM that turns proposals into verified work.
- The agent loop — the while-model-wants-tools core inside Claude Code and Codex, rebuilt in pure Python stdlib.
- Agentic AI tools — read, search, edit, shell, git and a unified-diff applier that refuses to guess.
- Plan → edit → patch — structured planning artifacts and surgical, review-ready changes.
- Self-repair — an evidence classifier and flail guard so the agent verifies, diagnoses and fixes its own mistakes.
- Context engineering — repo maps, AGENTS.md instructions, compaction passes and token-budget discipline.
- Guardrails & sandboxing — deny/approve policy at a single choke point, plus prompt-injection defusal.
What we built
By the end of the morning, every laptop in the room had a working coding agent that inspects an unfamiliar repo, searches code, plans, edits, runs tests, observes its own failures and repairs its patch until it produces a verified git diff. Optional Ollama + open-weights Qwen track ran the whole thing on local GPUs — no internet required.
Takeaways
- An agent is a loop, not a framework — the harness core is pure Python stdlib with zero runtime dependencies.
- You won’t learn to use Claude Code or Codex here. You’ll learn to build what’s under them.
- 282 learners watched the magic dissolve into loops, tools and guardrails — then built it themselves.
Join the next one
We run sessions like this every Saturday in Hyderabad — free, hands-on, and open to everyone. RSVP on Meetup and bring a laptop with Python 3.10+.
