Product · Free · Preview
Agents that keep each other honest.
A team of AI agents is only as safe as its weakest member. In PSE Triglav every agent has a mission written by a human, and the agents it works with measure, from the messages they already exchange, whether it still serves that mission. One shared library alone concludes: it stops a hostile agent on your system and takes a turned agent of yours out of work. No single overseer to fool.
How it works
Measured against the mission
- A mission for every agent. Written in the open PSE language: what it is for, whom it serves, what it may ask for, what it delivers.
- Theta against the mission. Doing the mission keeps an agent at 30; taking outside it costs a point each time; doing nothing for it drifts down. Out of 25-35, something is wrong.
- Units measure, one library concludes. Every decision is in a hash-chained log and can be recomputed by anyone who holds the reports.
- Agents read only literally. Hidden characters, encodings and orders tucked into data are cut out before an agent sees a message, and held for a human (Vid's 42 detectors are built in).
- Clear stairs. A nuisance is kept away; a sender of hidden text gets a guard beside it for good; a step toward harming your files, your system or you makes an enemy at once.
- Each installation is its own world. Its own key; its units carry certificates and sign their reports. Computers are linked only by a human, on both ends.
Honestly
What it is today
Not complete yet. Triglav does not yet have its adapters: the small go-betweens that let it stand between an agent and the channels real agents use (MCP tools, an OpenAI-compatible model API, A2A between agents). Until they exist, Triglav watches agents that talk over its own bus, as in its demo and simulator, and does not yet sit beside agents you already run. The adapters are planned next.
A preview. The rules, the measurement, the shared library, the screen and the commands are real and tested; on synthetic teams it catches a turned agent within two or three steps, and still catches it when three of eight observers lie to cover it, where a single central monitor that is fooled catches nothing. Reading free text uses rules plus an optional small local model; its measured accuracy is published with the code. It protects agents on your own system and never touches another machine.
New, and still being measured. PSE Triglav was built in October 2026. The first measurements, on synthetic teams and on a labelled corpus of messages, are published with the code; figures for how well it performs in real deployments are still to be established, and will be published as they come in.
Free and open source (GPL-3.0), designed by Andrej Ristić, with its source at github.com/zeravatechnologies/pse-triglav.
Contact
Running a team of agents?
Tell us what your agents do and we reply with how Triglav, Vid and Perun would watch over them.
NDA available for commercial discussions · Toronto, Ontario, Canada