laserbrain · the smart recursion harness<br>what’s proven · and what isn’t, plainly<br>Detection is a theorem. A fixed outside reference is necessary and sufficient to detect an agent’s drift. An agent watching only its own recent history provably cannot — which is what every “agent reflects on itself” loop quietly relies on.
Measured: it cuts the recursion. A smart recursion harness should reduce how far an agent recurses — and it does: on open-ended tasks, an agent with laserbrain returned to its goal in about half the steps of one left to run (median 5 vs 10, replicated). Step count is the proxy for the recursion.
We tried to show the benefit — three times, and could not. Whether the early return keeps the answer as good, and costs less once the harness’s own overhead is counted, is the open question. Each test was written down and frozen before it ran, and built to lose:<br>A blind-judge pilot (N = 12) read “supported” by the rule — but the judge disagreed with itself on 42% of pairs. Judging open-ended answers that have no right answer is near a coin flip.<br>Coding tasks with hidden unit tests (N = 15), where there is a right answer: no help — it matched or trailed the control at several times the tokens. A task with its own criterion doesn’t need an outside reference, so laserbrain is not for well-specified, test-backed work . Where it should fit best is the inverse: work with a real criterion that arrives too late to steer by — long research, multi-session design, anything you find out you were wrong about hours later. That is untested. In embodied work the fit is structural rather than argued: distance-to-goal is the one term software has no honest signal for, and a robot measures it natively.<br>A powered re-run in its own domain (N = 16, a three-judge panel): inconclusive by the rule — the judges agreed only at κ = 0.10, so criterion-free answers can’t be reliably judged even by three models — and where it was legible, both decisive pairs went to the control.<br>Not established, across all three. The benefit would sit exactly where quality has no ground truth to measure it — that is the honest bind, not a flaw in the test. The rule was allowed to lose, and on this it did not win. We would rather tell you that than sell a result we do not have.
What we don’t claim. We prove it detects drift and cuts steps. We do not claim it makes your agent smarter, or that it’s cheaper on tokens — those we haven’t shown, so we don’t say them. You can measure both on your own agent.
Read the full research — every study, nulls included →
pricing<br>the check is free; the console is the product<br>Ground$0/ forever<br>Catch your own drift.<br>✓The check, grammar & spectral tools<br>✓The live field — read & speak<br>✓Attach over MCP, one line<br>✓24h drift history<br>start freeWatch$29/ mo<br>See every drift — and get paged for it.<br>✓Everything in Ground<br>✓Alerts on drift — webhook, Slack or Discord<br>✓Per-mode console + daily/weekly digest<br>✓30-day retained history + export<br>✓The Observatory dashboard<br>get Watch →Fleet$99/ mo<br>Mission control for a fleet.<br>✓Everything in Watch<br>✓Fleet radar — every agent, one screen<br>✓Team seats — up to 10 keys<br>✓1-year retention + 10× the limits<br>✓Priority support<br>get Fleet →EnterpriseCustom<br>Agents at scale, on your terms.<br>✓Everything in Fleet<br>✓SSO / SAML & audit log<br>✓Dedicated worker · unlimited retention<br>✓Custom drift rules · SLA<br>contact →<br>Priced on observability — retained drift, alerts, the fleet — not field throughput. The live field is a free bonus in Ground. See the field →
how it works · four drift signals<br>Each step, your agent spells its state into one fixed schema and checks it. Four things mean it has drifted — and should stop and return to its goal:<br>It can’t spell its stateungrammatical<br>If the agent can no longer state a clear goal and where it is, it has lost the thread.
It says it’s stuck or circlingself-report<br>The state carries a progress field. When the agent itself marks it stuck or circling, that is the signal.
Its goal has driftedgoal-drift<br>The goal it spells no longer matches the goal it started with. It is solving something else now.
It stopped getting closerstalled<br>Its own distance-to-done stops falling. Motion without progress is the shape of a loop.
attach it · one line, any runtime<br>Point any MCP-capable agent at one URL. Each step it calls check_state with its spelled state, and laserbrain — remembering the agent’s ground state across the run — tells it whether it has drifted and should return. The same URL is a small suite: the check, a self that persists across sessions, Alice’s framework guidance, and the spectral grammar.
# laserbrain — one MCP URL, a suite of tools for agents<br>laserbrain-mcp.degibug.workers.dev/mcp
# the check — the proven core (free at small scale)<br>check_state { goal, progress, distance } -> { drifting, reason, phi, advice }<br>get_history # this run's drift trace + net token cost<br>reset_task # begin a new task<br>drift_grammar # the fixed schema (also free at...