Claude Code and AI agents for an engineering team

Omer Bin Saleem, a frontend chapter lead based in Riyadh, Saudi Arabia, runs Claude Code and a team of AI agents in his team’s daily engineering work. This is what runs, the numbers so far, and how to set up the same.

What runs every day

1. Code review596

   merge requests reviewed by an AI reviewer in 7 weeks, across 15 repositories.

   It takes the first pass on every merge request, so people spend their review time on design.

2. Team assistant101

   tools connect a Claude-powered assistant to Jira, GitLab, Slack, Jenkins and Confluence.

   It answers "is my fix deployed?" and "what is ready to merge?" in Slack, in seconds.

3. Agents17

   narrow AI agents with one lead. Each one has a single job.

   Pushes, merges and messages wait for a human yes. Everything else runs on its own.

4. Memory200

   plain files the agent reads at the start of every session. 85 of them are corrections.

   One fact per file with the reason, so a wrong rule is easy to find and delete.

5. Judge model245 ms

   median answer time of a small model that handles the many small yes or no questions around the agents.

   116,097 decisions in 20 days. It acts alone only above 0.9 confidence; below that, the old rule decides.

How to set it up

1. Write the rules where the agent reads them. A CLAUDE.md or AGENTS.md in the repo: conventions, test commands, and what never to touch.
2. Give it memory in plain files. One fact per file, with the reason, and an index the agent reads at the start of every session. A correction given once in chat is gone next session; written down, it holds.
3. Put AI on the first review pass. People review architecture and design. Track how long merge requests wait for a human, not just how many ship.
4. Gate everything irreversible. Merge, deploy, delete, pay and send wait for a human yes, on a device the agent cannot read. Tapapprove does this from an Apple Watch.
5. Route by risk. Small questions go to a small, fast model. Anything about merge, prod, deploy or release goes to the strong model.
6. Review instruction files like code. AGENTS.md and CLAUDE.md are instructions your agent will follow. Changes need a reviewer, and an agent never starts in a folder like Downloads.

What went wrong, and the fix

The small judge model once sent “merge this to PROD” to the small model, which missed that the work was already merged. Since then, merge, prod, deploy and release always go to the strong model.

Open source

• ClaudeTwin: two Claude accounts, one shared memory.
• Tapapprove: approve Claude Code’s permission prompts from an Apple Watch.

Questions people ask

How do you use Claude Code with a whole engineering team?

Rules in a file the agent reads, memory in plain files, AI on the first review pass, and a human yes before anything irreversible. The six steps are on this page.

How do you stop an AI agent from doing something irreversible?

Put merge, deploy, delete, pay and send behind an approval the agent cannot give itself, on a channel it cannot read, such as a button on a phone or a watch.

How do you give Claude Code memory between sessions?

A folder of small files, one fact each with the reason it matters, plus an index file the agent reads at the start of every session.

Who is Omer Bin Saleem?

A frontend chapter lead based in Riyadh, Saudi Arabia, who builds with AI every day and publishes one practical AI hack a day at omerbinsaleem.com.

Written by Omer Bin Saleem (Omer Saleem), frontend chapter lead, Riyadh. More: daily AI hacks, researched AI trends, LinkedIn.