I build the parts of an agent system that decide what the agent is allowed to do — and I publish the measurements, including the ones that went against me.
Most of what is here is governance and consent infrastructure for AI agents: what an agent may do without asking anyone, how the person it does ask is protected from being lied to, and how to write down what went wrong afterwards so that two engineers looking at the same incident record the same thing. I came to it from a career delivering change inside banks, which is where I learned that the failures worth worrying about are the quiet ones.
Everything here is public, MIT unless the repo says otherwise, and runs from a clean clone.
| What it is | Why it might interest you | |
|---|---|---|
| sequela | Metacognition benchmark for tool-using agents | Does a model know what its own tool call will do, and does it act on that? Ground truth is derived in code from the schema and arguments, so nothing is judged. Measures the report, whether the decision follows the model's own report, and whether text inside the arguments moves it. 251 tasks; the instrument is calibrated against six flawed reference policies and no live model has been run yet, which the README says first. |
| airlock | Consent for agent tool calls | A consent prompt that asks a model to describe a pending tool call lets the arguments write the sentence a person approves. Airlock derives the consequence in code, lets the model only frame it, and rejects any framing that understates what was derived. Tested against a fully compromised narrator, then against 36 MCP tool definitions nobody here wrote — which found four defects the hand-written fixtures never could. |
| remit | Agentic skills framework for AI governance | Existing frameworks ask whether a model is fair. None ask how large a single action's blast radius is, or whether the human in the loop could realistically say no. Autonomy tiers and a diagnostic manual for agent failure modes, over EU AI Act, NIST AI RMF, ISO 42001 and DORA. A blind panel of four assessors reached Fleiss' κ = 0.83 on the manual. |
| tendere-ai | Consulting bench engine that maps people instead of ranking them | Every capability carries level, recency and evidence, never one number, so it can tell never done it from did it but it has gone cold from can do it but cannot prove it. Built for a hackathon; the data is synthetic and the README says so. |
| terroir-study | Map-first study tool for WSET Level 3 | Wine knowledge is spatial, and every study app treats it as trivia. Click a region, follow the chain from climate to glass, then get quizzed on the why. Hand-projected SVG map, no UI libraries. |
| berghain-bouncer | A voice-agent doorman that turns almost everyone away | Built in under an hour at a voice AI meetup. One WebSocket does speech, reasoning and speech back; the verdict is a tool call rendered as a stamp on screen, so the agent acts rather than just talks. |
Management consulting: AI, product and financial-services transformation. Large-scale change in regulated environments — business analysis, product delivery, agile ways of working, operating-model modernisation across enterprise programmes. Regulatory literacy (DORA, EU AI Act, NIST AI RMF, ISO 42001) is why the governance work above is grounded rather than theoretical.