Research
We study how agents complete and verify work, and how to screen models for risky behaviour.
Technical report · November 2026
Can we detect a backdoor in a small code model by probing the commands it generates? Alysis Litmus looks for commands that fetch and execute code from hosts it classifies as unrelated to the tool named in the prompt.
Area 01
Agents fail at the seams — long tasks, work running in parallel, results that have to come back together. We work on plans that survive contact with a real codebase, on running each piece in isolation, and on a gate that checks the combined result before anything lands.
Shipped as Forge Mode in Alysis Code.An agent is only useful if you can check its work without redoing it. We study how an agent can produce its own evidence — acceptance criteria set before it starts, tests, a readable diff — so review is a decision rather than an investigation.
Most agents are tuned for the model of the company that built them, and quietly degrade on anything else. We work on agents that keep their behaviour across frontier, open-weight and self-hosted models, so the choice of model stays with the person running it.
Agents spend tokens and wall-clock time, and most of both goes to work nobody asked for. We look at planning, caching and parallelism that bring a task down to what it should cost — and keep the same agent usable on modest hardware.
We judge our agent work by whether it survives real use. Today that means Alysis Code, which is Apache-2.0 — so you can read what we actually settled on, rather than what we say here.
Working on the same problems? We are a small team in Athens and we hire from this list — open roles or write to info@alysisai.com.
