Deterministic gate
Plain-JSON rules block, hold for a signature, or allow each action before it runs. Speed limits trip a circuit breaker. No model in the decision.
AI can now refund customers, send emails and change records faster than anyone can check. Cosignary lets your team spot-check that work the way a good QA department would: it stops risky actions before they happen, picks a fair random sample for people to review, and gives you proof the work was right. No AI grading AI.
The demo runs in your browser with synthetic data: audit an AI support agent, then try to cheat the audit.
AI agents issue refunds, change orders, send emails and ship code. Two things make watching them hard.
Today's safeguards increasingly use one model to approve another model's actions. The watcher can be fooled, can drift, and shares the failure modes of what it watches. A note in the agent's own output can talk an AI reviewer into approving it.
Agents produce more actions than any team can review, and the gap widens with every model generation. "A human checks everything" was never true. Sampling at random is the only honest answer, and only if nobody can rig the sample.
Cosignary sits beside your AI agents, scripts and automations. It gates what they're about to do, records what they did, and turns an endless stream of actions into a few questions a person can answer.
Your systems
Plain-JSON rules block, hold for a signature, or allow each action before it runs. Speed limits trip a circuit breaker. No model in the decision.
Every action goes on a hash-chained ledger. A seed locked in beforehand picks the sample, and anyone can recompute it afterward.
Your own people answer one short question per card. Planted test cards measure whether they're really looking.
Works on AI agents, ordinary scripts and human workflows alike: an action is just a JSON record.
Nobody can read everything an AI produces. A random sample nobody can game, checked carefully, says something precise about all of it.
The 299 holds at 95% confidence whether a category has ten thousand actions or ten million, as long as none of the checks finds a problem. Review time plans on 350 checks per category a month, so a few problems can turn up and every category still proves under 1% each quarter, at a careful two minutes per card. The demo's five attempts: editing a past action, rebuilding the chain, deleting an action, cherry-picking the sample and swapping the seed.
Every action passes the gate first (human-in-the-loop, designed for agent volume). Most run straight away. The risky ones wait for a person's signature, and a burst of small ones trips a circuit breaker, because per-action caps can be beaten by volume.
Order total $640. Over the $200 cap.
21 refunds of $24 in 10 minutes. Each one passed the $200 cap.
3 problems in 52 checks; audit rate doubles tomorrow.
"More than 20 refunds or $2,000 in 10 minutes." Counted across actions, so 300 small refunds can't slip past a per-refund cap.
The ledger proves who approved, not just that someone clicked. Flip a signed "deny" to "approve" and verification fails.
Held actions expire to deny. Irreversible actions run only at a pace people can actually sign for.
Now: approvals, tripped breakers, tamper. Daily digest: categories above target, missed planted tests. Webhooks reach Slack, Teams, SMS and pagers.
The election-audit pattern (a seed committed in advance, a selection anyone can recompute) applied to automated work.
Publish only the hash of a secret seed, before anything happens.
Every action is chained to the one before it, then checkpointed outside.
Check an action when HMAC(seed, fingerprint) falls under its rate.
Yes, No or Not sure on each card, with planted tests mixed in.
Publish the seed and the planted-card key.
Recompute the sample. Any edit, deletion or cherry-pick fails.
Rates adjust by a published rule. If a category's problem rate comes in above its target, its audit rate doubles next period, up to 100%. A clean period halves it back toward the base rate. Anyone can recompute that from the ledger too.
The cloud part of Cosignary is a witness, not a judge. It holds no customer data, no seeds and no decision power, so breaking into it can't rig an audit.
A witness stores hashes and counts. An attacker who breaks in learns nothing about your business.
Each witness signs its own chained log, and you keep the receipts. Backdating or forging a checkpoint breaks a signature.
The worst a breached witness can do is go silent, which shows up as missing receipts. Anchor with several.
Every report states its level plainly, so a self-reviewed period never reads as an independent one.
Your own people review the sample. The selection is still fair, but only you vouch for it.
For small businesses and internal assurance. Free and open source.
The seed is revealed and the public checks pass. Anyone can confirm the sampling was fair without seeing your data.
For showing customers and partners the process is honest.
An outside reviewer re-judges a seed-picked share of cards, redacted and in your environment. The stricter verdict counts, and agreement is reported.
For insurers, enterprise buyers, regulators and auditors.
The same protocol fits every industry. What changes is the policy, the checklist on each card, and who the watchers are: in each case, the people who used to do the work.
Nobody buys trust in the abstract. They pay when its absence is costing them a deal, a discount or a compliance finding. And because review is sampled, the cost stays small.
Hours a month to prove every category is under a target problem rate, at 95% confidence.
Watcher (n.): the person accountable for confirming that automated work in their field is correct, judged against recorded evidence, and for signing off on anything that can't be undone.
A fitter checks fittings, a bookkeeper checks invoices, a support lead checks refunds. The credential is a live, measured catch rate on planted tests, renewed from evidence rather than earned once in an exam, so pay can follow accuracy instead of speed.
What jobs AI will replace and what's left, the six-level watcher career ladder, roles by industry, and how watchers are measured.
More in the human-in-the-loop guide, the watcher guide and the security model.
Cosignary is an oversight layer for AI agents and any software that takes actions on its own. Deterministic rules decide whether each action runs, a tamper-evident ledger records everything, verifiable random audits choose which actions people check, and planted test cards measure whether the checkers are paying attention. No AI model makes the decision.
Most agent monitors use an AI model to judge whether another model's action is safe. Cosignary keeps AI out of the decision: rules and code decide, people sign, and the evidence (ledger, sample, signatures) can be verified by anyone. An AI may flag an action for extra review, but it can never clear one.
No. Selection, enforcement and verification are deterministic code, and people give the verdicts. AI can help, but only by adding scrutiny: it may flag an action for review, never clear one. An AI flagger must come from a different vendor than the agent it watches, and it's scored on planted tests like any other reviewer.
You can build the checker. You can't build someone else's signature on the result. A self-audit isn't independent; a homemade method has to be defended to every auditor and insurer; and benchmarks, planted-test libraries and certified reviewers only exist across many organizations. The core is open source on purpose, so building your own gains you nothing.
Yes. Every action passes the gate before it runs: allow, hold for a signature, or block. Speed limits trip a circuit breaker that pauses a whole category, such as all refunds, until a named person resets it. Held actions that nobody answers are denied, never approved by default.
No. Everything that touches data runs in your environment. Only fingerprints (hashes and counts) go to outside witnesses. Review cards are redacted, and outside reviewers, if you use them, work through your own secure access.
No. Self-review is the default and a first-class mode: your own people, your own servers, no outsider. Turn on outside review the day an insurer, an enterprise customer or a regulator asks for evidence.
Review grows with the confidence you want, not with volume. Plan on about 350 random checks per category each month. That proves the category is under 3% every month and under 1% every quarter, even when reviewers find the handful of problems a good agent produces; 299 clean checks prove under 1% in a month only if they find none. A million actions across five categories is about 1,750 cards a month. At a careful two minutes a card, that's roughly 60 hours: a part-time job for one reviewer, whatever the volume.
Yes. Any system that takes actions can log them: scripts, scheduled jobs, code commits, device-fleet updates, or people using internal tools.
Built by a founder who runs AI agents in production. Sean Morrissey built the guardrails behind Cosignary to keep AI agents from breaking his own live product, FitMyGolfClubs. Read the story.
Review a verifiable sample, catch the planted tests, try five ways to cheat the audit, then approve a held refund from a phone and trip the circuit breaker.