ATLAS

The layer

Observe → Predict → Decide → Act → Govern. ATLAS is the whole loop, which is rather the point of calling it an operating system. It observes what agents actually do, predicts what a piece of work will take, decides which model earns it, drives the run, and governs the money and the outcome — then feeds the result back into what is allowed to run tomorrow.

Daughter of The Governor

The Governor is the idea, published in the open and kept deliberately educational — the argument that when programs became untrusted, computing grew an operating system, and that agents are at exactly that moment now.

ATLAS is what the idea grew into. Same core, same one architectural rule, but packaged, front-doored, audited and pointed at real money. The Governor explains why the layer must exist. ATLAS is the layer, installed on your own hardware, with a console you can watch a run die in.

The two questions

Everything else in the agent stack of 2026 already exists — runtimes that drive tool calls, orchestrators that queue work, billing systems with the right schema. What existed nowhere is something that can answer:

Did the agent actually do the work it claims it did? What did that cost, and was it allowed to spend it?

Every run goes through the same governed lifecycle, and there is no fast path around it:

select → guard → reserve → checkpoint → run → verify → contain → settle → record → earn trust

The loop closes. What a run did becomes what the system is allowed to do next — and that closing is the entire difference between an agent harness and an operating system.

What it refuses to do

A governor is defined rather more by what it won’t let you do than by what it will, so here is the list of arguments ATLAS will lose on purpose.

It will not invent a budget. There is no default. A governor that guesses a spending limit on your behalf has comprehensively missed its own point.

It will not run something it cannot verify. No check and no judge means the run is refused before reserving a penny — paying for a run whose outcome was decided before it started is precisely the waste this exists to prevent.

It will not let a good average excuse a failure. One verified failure drops the trust tier to zero. Averages are how you talk yourself into things.

It will not price a model it does not know. An unpriceable model is one the lease cannot bound, so the run does not start. null is not zero, and pretending otherwise is how you find out what your credit limit was.

It will not grant trust. Not from the CLI, not from the console, not because you are the owner and you are quite sure it will be fine.

Where it sits in the stack

ATLAS is the layer agents run on. Saksi is the memory they run with — the provenance-backed record of what was actually said and promised — and SpeakEasy is how the human end gets into that record without typing. Voice in, verified memory in the middle, governed execution on top, and a confirmation gate between the last two.

An agent is only ever as safe as what it knows and as affordable as what it is allowed to spend. That is two problems, and they want two products.

Status

Private beta. 620 automated tests, green and CI-gated on every push; aos run drives a real agent against a real budget and settles at the true cost whatever happens. Production integrity, security and capability audits are on file, and the sweep of 6 August 2026 confirmed the thing is genuinely built and wired end to end — no stubs, every front door reaching the real governor. It also found real defects in three seams, which is what an honest audit is for and why this says beta rather than something braver.

Curious about the reasoning rather than the runbook? Start with The Governor — why the agentic age needs an operating system, or request a briefing.