State lives in plain files.
A task is a markdown file in a folder, and the folder is the state. No hidden database, no dashboard that lies. Anyone — human or agent — can read the truth with ls.
AI Engineering Consultancy · Sheffield, UK · Remote
Systems that
close their own loop.
Closed Loop designs and builds production AI systems for businesses that need them to be right: autonomous dev loops on Claude Code, multi-agent pipelines, the cloud infrastructure underneath them, and the security audits that prove what AI built is safe to ship. Every method we sell ran for real before it was written up. No thought experiments —scars with documentation.
Most AI demos are open loops: a prompt goes in, text comes out, and nobody checks. A closed loop feeds its own output back through a gate — verified, reviewed, recorded — before it's allowed to count. That gate is the whole discipline, and it is what Closed Loop builds.
How we work
Three rules, each one a scar.
Verification by evidence, not assertion.
An agent saying "done" is a claim. A baseline diff, a test run, a screenshot in the record — that is evidence. Every loop we build has a gate that only evidence opens.
Separation of duties, enforced by capability.
The agent that writes code cannot merge it. The agent that reviews cannot edit. Not because the prompt says so — because the tools it holds make it impossible.
Latest writing
Systems, explained from the inside.
Available for remote engagements
Need an AI system that has to be right,
not just impressive?
Agentic system design, Claude Code enablement for engineering teams, the infrastructure that keeps LLM products production-grade, and security audits of AI-built and AI-powered systems — delivered by a senior engineer, not a slide deck.