Coding and design agents
Haris Muhibić built agents that turn bug tickets into a reproducing test and a draft pull request, answer code review comments, and check a designer's Figma frames against what the codebase can actually build.
- Role
- Sole author
- When
- 2026
- Stack
Ticket to draft PR
A bug labelled for the agent is triaged, routed to the right repository, reproduced with a failing test, fixed and opened as a draft pull request.
- The agent never pushes. It only edits files and runs tests. A deterministic pipeline step creates the branch, commits, pushes and opens the PR against the correct repository.
- Triage runs first. Verdicts are fix, needs more info, or won’t auto-fix. Auth, payments, schema changes, cross-service changes and large diffs are always excluded.
- Reproduce first. Front-end bugs get a Playwright test; back-end bugs get a Jest test in the target repository’s CI.
- No loops. A hidden marker plus the bot’s commit identity let the agent tell its own activity from a human’s, without storing any state.
- Cost on the ticket. Each run posts its cost back to the ticket.
I first built a custom harness, then deliberately replaced it with an off-the-shelf GitHub Action. That is a build-versus-buy story worth telling.
Design buildability agent
A Claude Code workspace for a product designer. It reads live Figma frames and answers in plain language whether a design is buildable today, buildable with new work, or not supported. It checks design tokens, components, back-end endpoints and platform limits.
- Clear source of truth: Figma variables win on colour; code wins on behaviour.
- Fresh knowledge: a headless job rebuilds the code-derived knowledge files every night, and a session hook updates the workspace when the designer opens it.
- Gated writes: writing to Figma stays switched off until a one-time write test passes.
Building something similar? Book a 30-minute call or email haris@chern0.dev.