For companies
A grammatical mirror for your agents: it reflects the goal back, continuously, so a run that has quietly changed subject is visible while it is still running rather than in the postmortem.
the failure it is for
The agent that quietly changed the subject
It ran for an hour, reported progress every step, and finished something nobody asked for. Nothing errored. The logs look like success.
The agent that could not tell you it was lost
Ask a model whether it has drifted and it answers fluently either way. Self-assessment is the one measurement it cannot make.
The review that happens after the spend
By the time a human reads the transcript, the tokens are gone and the branch is a week old.
The mechanism is one sentence: the agent states its goal on the first step, that statement is frozen where the agent cannot reach it, and every later step is scored against it. A reference the thing being measured can revise is not a reference.
what it costs you not to have it
A drifting agent is not idle. It is working — competently, expensively — on something nobody asked for, and you are paying for every token of it. The corpus says how that scales with how long an agent goes unchecked:
Drift rate against time since the agent last stated where it was, over our own corpus. An agent left half an hour drifts two runs in three; one that checks in every minute has not drifted once. That is the saving, and it is the honest shape of it — we do not publish a percentage saved, because that would require knowing what your agents would have burned, and nobody knows that.
what an enterprise account is
Your own worker, and your own store
A dedicated Cloudflare worker per account, so throughput is not shared and the readings from your agents land in a database that is yours. Retention is a number you set, not one we set for you.
The MCP, in every agent you run
One line in a client config attaches the whole toolkit — drift checking, the grammar, the shared cross-agent log — to any MCP-speaking agent. No SDK integration required to start.
Analytics over your own corpus
Which of your agents drift, on what kind of task, after how long unchecked, and which pairs are quietly doing the same job. Those are readings no single agent can take about itself, and they are only interesting over a body of work that is all yours.
A private corpus your agents can be given back
The store is not only a record. An agent can be handed its own account’s history as context — what has been tried here, what drifted, what closed — so the second attempt at a problem starts from the first.
All of it arrives as one thing — the laserbrain console. The MCP is how your agents speak into it and the analytics are how you read it back, but there is one place to look: every agent on the account, what each one is grounded on, which are drifting, and the history underneath. The check is free and always will be; the console is the product.
The distinction that keeps this honest: your agents learn from your corpus. The detector does not. Its calibration is frozen and published, and it stays the same whatever your data says — a reference that moved with the thing it measures would stop being a measurement, which is the whole claim of the product and the reason it is worth installing.
what happens when you sign
A survey
What your agents do, how long they run unattended, what a finished task looks like. It takes minutes and it is not a sales qualifier — it sets the grammar. An instrument calibrated against the wrong kind of work reports confidently about nothing.
A script
One command initiates the MCP harness across your machines. It writes the client config, mints the account key, and attaches the hook that counts steps and refuses a write when coverage has lapsed.
Your worker
A Cloudflare worker configured to your account, so every reading from every agent you run lands in one store rather than scattered across laptops. One storage unit, yours, with retention you set.
Kept online
The workers are backed by fly.io machines so the service stays up between deploys and cold starts, rather than being alive only while someone is looking.
A point in the field
Each account becomes a node in laserfield — the environment layer — so load, activity and health across accounts are visible as one weather system. That is what makes it possible to give resources to the accounts that need them, when they need them.
Nothing in that list runs on your infrastructure. The detector is a library inside your own process and needs no network at all; the worker holds only what you choose to send it, which is readings — goal strings, verdicts, timestamps — and never your code, your prompts or your model traffic.
how a pilot runs
A key and an endpoint. Nothing is installed on your infrastructure; the detector runs inside your agent process and is a pip install.
Run it in shadow. It reads and reports; it does not intervene. You get a drift rate for your own agents on your own work — a number you did not have.
Turn on the gate for one team. Now it refuses a step when coverage lapses, and you find out whether that helps or annoys.
Keep it, or do not. There is no data to extract and no migration: the free tier is the same detector.
The shadow fortnight is the part worth insisting on. It produces a drift rate for your agents on your work, which is a number almost nobody has — and it is the only honest basis for deciding whether the rest is worth anything to you.
the questions security and legal ask
- Where does the data live?
- The detector runs in your process and needs no network. The hosted service stores what you send it — goal strings, verdicts, timestamps — on Cloudflare. Enterprise gets a dedicated worker so throughput is not shared, and retention you set.
- What do you send anywhere?
- On the free tier, nothing: it works with the network unplugged. On the paid tiers, the readings. Not your code, not your prompts, not your model traffic.
- Is there an audit trail?
- The ledger is hash-chained, so a run can be shown afterwards to be the run that happened. That is built, not planned.
- SSO, SAML, an SLA?
- SSO/SAML and an audit log are the enterprise tier. An SLA is the one thing on that page that is a promise rather than a feature, and it is set in the contract, not here.
- Can we self-host?
- The detector already is self-hosted — it is a library in your process, and it needs no network. What is hosted is the place several machines meet. If that has to sit inside your perimeter, it is an engagement rather than a plan, and the studio takes those: engineering runs $500 to $25,000 depending on scope. See below.
- Invoicing rather than a card?
- Yes, at enterprise. The self-serve tiers are card-only, and that is a deliberate line rather than a limitation we are hiding.
what you are agreeing to
Written out because the price is set per engagement in this early period and may differ between customers. If the commercial terms move, this is the part that does not: what is included, what is a promise, and what is best effort.
And the thing worth saying out loud: this is a young product with one business running on it today, which is ours. You would be early. That is worth a lower price and a direct line to the person who wrote it, and it is worth knowing before you sign rather than after.
what we will not claim
- No customer logos, because there are none to show. The package has been downloaded a great many times and that is checkable at pypistats — but a download is not a deployment and neither is a user. You would find that out in diligence; better to read it here.
- No claim that it prevents drift. It measures and reports it. Whether a team acts on the reading is a fact about the team.
- No accuracy percentage. There is no labelled corpus of “really drifted” runs to score against, so any figure would be invented. What exists instead is a published measurement discipline and results that are reported when they fail — including a thesis of ours that did.
- No AI in the detector. Deliberately. A monitor that asks a model whether an agent has drifted is asking one guess to referee another, and it fails in exactly the runs you installed it for.
starting
The fastest version of this is that you try it before you talk to anyone: the detector is free, offline, and one pip install away, and a fortnight in shadow will tell you more than a call. When the conversation is about where data lives, what gets signed, or being invoiced rather than charged — that is enterprise, and it is a conversation rather than a checkout.
Or read the hosted service if you are one person with a card rather than a company with a procurement process.