Build a multi-agent dev platform in three weeks with spec-first agent skills

Essaïd LaBani pair-programmed a self-hosted multi-agent developer platform with an AI coding agent: 173 commits, 66 planned phases and about 1,200 Python tests between 21 October and 10 November 2026.

Done withCustom agentinGitHub· Nov 15, 2026

What they did

Essaïd LaBani built Agentic Dev Platform, an open-source tool that runs multi-agent workflows against sandboxed repository clones and opens pull requests after a human approves. Most of the work was pair-programmed with an AI coding agent the post does not name. Four spec files (mission, roadmap, tech stack, design language) were set to outrank the code, the agent's instruction file and anything said in chat. Two skills checked into the repo drove each roadmap phase: a plan skill that found the next unplanned phase, opened a branch, asked the developer three multiple-choice questions on scope, approach and validation, then wrote requirements, a plan and a validation checklist; and an open-pr skill that wrote a changelog entry, opened the pull request and watched CI. A human squash-merged each one.

How it went

Between 21 October and 10 November 2026 the repo reached 173 commits, 66 planned phases, roughly 26,000 lines of Python, 28,000 lines of Python tests and 45,000 lines of TypeScript. CI rejected pull requests that skipped the changelog, left a finished phase unmarked in the roadmap, or let generated schema and API client files go stale. The platform is still pre-1.0 with milestones outstanding. Green tests against a fake model runtime hid three bugs that only showed up in real runs with Claude, which led to a lessons file of rules and the checks that enforce them.

Worth knowing

The coding agent used to build it is not named, and the numbers come from the author's own write-up. Tests with scripted fake agents only proved the workflow shape: a lead agent with no tools, a reviewer reading the wrong branch and a synthesizer with no input all passed them. The author now reads a real run step by step before calling a phase done.

Try it

With Custom agent· 3 parts to fill in

We build [project] spec-first. Treat the files in [specs folder] as the source of truth over the code and over anything I say in chat. Find the next unplanned phase in [roadmap file] and open a planning branch. Ask me three multiple-choice questions: scope, approach and how we will validate it. Then write requirements, a plan and a validation checklist into a dated folder, turning the phase's done-when line into concrete checks. Pause for my approval before writing code. Implement against the plan, add a changelog entry, mark the phase done in the roadmap, open a pull request and watch CI. Stop when every check is green and leave the merge to me.

Square brackets are the parts to fill in. The prompt is ours; the story is theirs.

Where they said it

Related