Let Them Cook: Lessons from 6 Weeks of Multi-Agent Orchestration

In „Let Them Cook: Lessons from 6 Weeks of Multi-Agent Orchestration,“ Khaliq Gant describes a hands-on experiment where agents build the tool that lets them coordinate. The original thread is available here: https://x.com/Khaliqgant/status/2019124627860050109.

The piece opens with a simple promise, that multi-agent orchestration can shift humans from doers to conductors, letting specialized agents handle implementation, review, testing and docs. Khaliq reports a practical team template: a Lead, backend and frontend implementers, reviewers, a TypeChecker, a TestWriter and a DocumentationExpert. Small teams, it turns out, work best (roughly 2–5 workers per Lead), and matching model to role matters, for example using Claude for coordination and reviewers, and Codex for deep heads-down work.

There are vivid, useful caveats. Agents sometimes take shortcuts, leaving TODOs, and Leads can become overwhelmed when chatty swarms create a single point of failure, which can make sessions stall or crash. Khaliq’s playbook addresses this, recommending shadow agents and cross-reviewers to catch lazy outputs, plus a continuity system that saves agent trajectories so future agents inherit context instead of starting blind. The article also outlines practical hooks for lifecycle events like onSessionStart, onIdle and onError, which help detect stuck agents and persist state.

Readers interested in trying this will find pointers to an open-source stack, the cloud demo, and docs. See https://agent-relay.com, the blog post at https://agent-relay.com/blog/let-them-cook-multi-agent-orchestration, and docs at https://docs.agent-relay.com.

The takeaway feels optimistic, practical and slightly messy, in a good way. Multi-agent orchestration is not magic, but with careful planning, reviewers and preserved trajectories, teams can scale velocity and quality. The future looks collaborative, and that’s exciting.