How I think about multi-agent systems2 min
AI · Multi-Agent Systems
When one agent isn't enough
A single AI assistant is a generalist with no checks and balances. Multi-agent systems borrow the oldest trick in engineering organizations: specialization plus review. One agent architects, another implements, a third tries to break it.
This hub covers what I've learned building an AI 'engineering org': what genuinely improves output, what's expensive theater, and where the hard problems hide (spoiler: evaluation and independence).
The essentials
Specialize by charter, not personality
Agents need narrow, testable mandates — 'review this diff for money-safety' beats 'you are a meticulous security expert'.
Make review adversarial
Agents agree with each other too easily. A finding should survive an agent whose only job is to refute it.
Orchestrate deterministically
Use code for control flow (loops, gates, fan-out) and models for judgment. Letting models improvise the pipeline compounds errors.
Latest on this
All writing →The Build Log
The stuff I learn while building.
I test a lot of tools, workflows, and ideas. The Build Log is where I share what actually worked, what surprised me, and what I would do differently.
AI workflows. Claude. Agents. Building software. Building companies.
- Practical AI workflows I actually use
- Tools and systems worth your attention
- Lessons from building software and companies
- Things I learned the hard way
One email when there is something worth sharing. From Charles, no one else.
