Issam Maher R.
All writing

How I engineer with AI agents: Claude Code, MCP and multi-agent workflows

The setup and habits that make AI agents useful on real production codebases, not just demos.

Issam Maher R. · 2 min read

I use AI agents every day across .NET, Next.js and Flutter codebases. The gains are real, but they come from the setup around the model, not just from the model.

1. Give the agent the project's memory

An agent starts every session knowing nothing about your decisions. I keep that knowledge written down where it can read it: the stack, the conventions, the things we deliberately don't use and why. ADRs do double duty here. They explain decisions to humans and stop the agent from "helpfully" adding a dependency we removed on purpose.

2. Connect it to real tools with MCP

The Model Context Protocol lets an agent work with the same systems I use: issues in Linear, deployments and logs in Vercel, errors in Sentry. "Fix the bug in this ticket" becomes a loop the agent can actually close: read the issue, find the error, change the code, check the deployment.

3. Split big work across multiple agents

For larger tasks I run a workflow instead of one long conversation:

  1. Research agents fan out to map the relevant code and constraints.
  2. Implementation happens with that context gathered.
  3. Independent review agents check the result, each from a different angle (correctness, security, simplicity).
  4. Verification comes last: a finding only counts once something has confirmed it, for example a failing test or a reproduced bug.

The review step matters most. A model reviewing its own work tends to approve it. A separate agent told to break it does not.

4. Keep a human on the irreversible steps

Agents draft, build and test freely. Pushing to main, deploying, deleting data and messaging clients still need my approval. That boundary is what makes it safe to give agents real access everywhere else.

What actually changed

The biggest change isn't typing speed. Work I used to put off now gets done: the second test case, the ADR, the migration note, the cleanup after the fix. The bar for quality went up because doing things properly got cheaper.