Looped systems · Kist

The open shell that runs the loop.

Kist is the agent shell at the core: bring a key for any model, hand it a goal, and it works to completion — assessing its own effort each cycle, verifying its own output, and convening a council before it calls anything done.

Summary

Kist is the open-source engine Perslis is built on. It is model-agnostic by construction — you bring the key, it drives the model — and it is a loop, not a single call: it self-assesses how much effort a goal needs, works a cycle, verifies the result, learns from its own history, and keeps going until the work is done and proven or a safety cap stops it. For anything high-stakes it convenes a council of models to review before it reports back. The shell, the safety gate, and the verification machinery ship as one checksum-verified binary.

The great-loop

A single prompt-and-response is a guess. Kist runs a loop that closes on evidence:

The council of minds

Before Kist trusts an important result, it puts it in front of more than one model. A builder drafts; reviewers from different labs critique; an adversary tries to break it; the strongest available model signs off. Disagreement between models is treated as signal, not noise — it is where real errors surface. See Models for who sits on the council and why.

Bring-your-own-key, open source, no lock-in. Kist never resells inference. Point it at the accounts you already have — or at models on your own hardware — and it does the rest. The value is the loop, the council, and the proof.

Where it runs

Kist runs the same loop against the cloud frontier or entirely offline. For sensitive work it drops into the AirTight mode — build, review, and record all local, nothing leaving — and for regulated domains it applies the citation-and-verification discipline of the Legal loop. One engine, tuned to the stakes.