Most "AI agents" stop at the first obstacle and hand the work back to you. I build the other kind: a system that keeps going until it either succeeds or has genuinely run out of options.

What I build

· An orchestrator plus specialised sub-agents for research, execution and verification

· A persistent state machine — search, evaluate, try, verify, adapt, retry — instead of a single prompt chain

· Shared memory between agents, so context is not rebuilt from zero on every run

· A verification layer: results are checked before they reach you

· Full run logging, so you can see what was tried, what failed and why

· Deployed on your own infrastructure. Yours to keep and extend.

I work with OpenAI, Anthropic, Mistral and open models, in Python and TypeScript.

We start with a scoping call. If the task genuinely does not need an agent system, I will tell you that and we stop there — I would rather lose the project than build you something that looks impressive and does nothing.

Typical build: 3-4 weeks for design plus a working first version.