AI coding agents specialist: Claude Code, Codex and vibe coding that holds up in production

I set up AI coding agents for engineering teams so the code they write can be trusted. Anastasiia Butova, ML engineer in Belgrade, Serbia. I build image-processing software with Claude Code and Codex every day, and I open-sourced the tooling that made that work in a team: rulebooks the agents follow, hooks that enforce them, coordination between parallel sessions, and a check that stops an agent from saying "done" while two languages disagree. Open to contract work and consulting next to my main line, ComfyUI and image-processing pipelines.

What a team gets

1. An audit of how agents work on your codebase. Where they guess, where a green test hides a broken seam, which instructions they ignore under a long context.
2. Rules that hold. CLAUDE.md and AGENTS.md written for the codebase, and the hooks that make the important ones mechanical: a rule in a prompt is advice, a hook is a gate.
3. Seams that are checked. Crossweft maps the hand-written contracts between Go, TypeScript and Python (an API version, a DTO field, an env var) and fails the change when one side moves alone.
4. Parallel sessions without collisions. mclaude: locks, handoffs and a mailbox for several Claude Code sessions on one repository.
5. Vibe coding with a floor. Prototype fast, but nothing ships on the agent's word: each change carries a test or a runtime check, and an independent review for anything irreversible.

Evidence

• codex-claude-code-config - 30 architectural principles, 58 skills and 40+ safety hooks for Claude Code and Codex; MIT, 150★ and 24 forks on GitHub.
• Crossweft - open-source cross-language contract check with an MCP server, agent hooks and a GitHub Action (source).
• mclaude - file-based coordination for parallel Claude Code sessions, 193 tests.
• Price index - agent-run market pricing for a jewellery wholesaler: four agent lanes with explicit allow and forbid lists, where no agent may write a price and only deterministic validators do.
• Glam.ai - built the multi-agent coordination system the engineering team used, an AI knowledge base (550+ articles) and pre-configured VMs with AI tooling for engineers.
• happyin.space - 1,287+ reference articles structured for agents to retrieve from.

Writing on agents

• My CLAUDE.md is 582 lines. Six layers, and the failure behind each
• Several Claudes on one repo. Locks, handoffs and email from 1982
• Every hand-written seam needs a guard
• 785 articles for agents, not humans
• Where Claude Code desktop sessions go when you switch accounts

Questions

Can you set up Claude Code or Codex for our team? Yes: an audit of the codebase and how the team uses agents, then the rulebook, the hooks and the checks, handed over in your repository.

Is vibe-coded code safe to ship? Not on the agent's word. It becomes shippable when every change is verified by something other than the agent that wrote it: a test that fails without the change, a runtime check, a contract check across languages.

Do you work remotely? Yes, from Belgrade (CET), in English, Russian or Serbian.

To start: one paragraph on your codebase and what goes wrong with agents today, by email or Telegram.

Elsewhere on this site

• Home - ComfyUI and turnkey image-processing pipelines
• CV - roles, dates, stack
• Blog - essays on agents, diffusion and production ML
• AGENTS.md - the machine-readable profile