I work on projects to learn how AI agents behave: what an agent is allowed to do, how a person can check what it's about to do, what it's allowed to remember, and how to measure whether it knows.
Everything here is MIT-licensed and most of it has tests that run on every push. The READMEs say what doesn't work yet: the benchmark below has only been run on two small models, not a large hosted one, and the Magic 8 Ball is safer than asking the model directly but not smarter.
| What it does | Why it matters | |
|---|---|---|
| sequela | Tests whether an AI agent can say what its own next action will do, and then acts on what it said. | The right answer is worked out from the tool itself, so there is no judge to argue with. Two small models have been run on a laptop. Neither did better than a few lines of rules, and the bigger one played danger down more often, not less. No large hosted model has been run yet. |
| airlock | Shows a person, in plain words, what an agent's tool call will do before it runs. | The words are built from the facts of the call, not written by the agent, so the agent cannot talk anyone into approving it. Tested against 36 real tools nobody here wrote, which found four defects the home-made tests never could. Attacking its own proxy later found calls that got through unjudged when batched or malformed; those are fixed, and the README says so. |
| keepsake | Lets a person approve what an AI agent remembers between sessions before it takes effect. | A note in memory runs in every session after it, including the ones nobody is watching. It keeps a copy of what you approved, shows each change in plain words, and holds an agent's write until you say yes. 20 of 21 planted changes shaped like this month's documented attacks score high or critical, and none of 21 ordinary notes do. A harmful line worded outside its vocabulary scores low, and the README says so. |
| remit | A rulebook for what an AI agent may do without asking a person, mapped to the EU AI Act, NIST AI RMF, ISO 42001 and DORA. | Two engineers looking at the same failure write it up the same way: four independent assessors agreed at 0.83, where 0.61 counts as good. |
| tendere-ai | Matches consultants to work by what they can do, how recently, and what proves it, instead of one score. | It can tell never done it from did it years ago from can do it but cannot prove it, and shows leaders which scarce skills people have but cannot yet prove. Built for a hackathon on made-up data, and the README says so. |
| eightball | A Magic 8 Ball that answers from a small AI model on your own machine, and only ever with the 20 classic phrases. | It reads the model's odds rather than its words, so it can hand back "reply hazy" when it is not sure — safer, not smarter: it speaks far less often and is wrong far less often when it does. |
| terroir-study | A map for studying wine, for the WSET Level 3 exam: nine countries and 77 regions. | Click a region, zoom in to its villages, follow the chain from climate to what is in the glass, then get quizzed on the why. 849 questions that remember what you got wrong, and a tasting drill marked against the grape. |
| berghain-bouncer | A voice doorman for a Berlin techno club that turns almost everyone away. | Built in under an hour at a meetup. The verdict is stamped on screen, so the agent acts rather than just talks. |