Prove what your AI agents did.
Sigilla is the system of record for AI agents. Ask what an agent did and get an answer you can prove: what it was asked, what it decided, what it said, and what actually changed in your systems.
Fig. 0 · The Sigilla console. Open an episode to see how it was answered. Example data.
Agents act on real money and real customers. Their logs can’t prove what happened.
The agent’s word isn’t the world’s word
An agent can say it issued a refund that never happened. Sigilla checks every claim against the system that actually moved the money.
“No record” isn’t proof
Missing logs could mean nothing happened, or that nobody was watching. Sigilla tracks coverage, so it can tell the two apart.
Logs aren’t evidence
Auditors, insurers and counsel need records they can check themselves. Sigilla hands them a package they can verify offline.
How it answers
Three answers. Never a guess.
Every question gets one of three answers, decided by deterministic code. No model decides a finding.
Found, with the records
The action happened. Here are the sealed records that show it, from the agent and from the system of record.
Provably didn’t happen
Not found, and every source that would have recorded it was provably watching the whole window.
Honest about the gap
Not found, but a source wasn’t watching for part of the window. Sigilla names the gap instead of guessing.
Swipe to see all three →
When it matters
For the moment someone asks.
Disputes and chargebacks
A customer says they were never refunded. Show what the agent did, what it said, and what Stripe recorded.
Audits
Answer control questions across every agent action, with evidence your auditor can check offline.
Insurance
Show insurers how your agents behave, including proof of what they didn’t do.
Incident reviews
Rebuild an agent incident from sealed records, and see exactly which sources missed what.
Swipe for more →
Fig. 1 · Coverage · example data
Know what you didn’t see.
Sigilla knows which sources should have recorded what, and when each one was watching. Every gap goes into a register, so “no record” is never mistaken for “didn’t happen”.
$ sigilla verify evidence-4821.zip chain intact manifest signature valid timestamp valid coverage: 1 gap, named verification-ready
Fig. 2 · Evidence packages · example data
Evidence anyone can check.
Export any answer as a sealed package. Auditors, insurers and counsel check it with an open-source verifier, offline, without an account and without trusting us.
From first trace to first answer.
- 01 / CapturePoint your agents’ OpenTelemetry at the Sigilla collector. It runs in your environment and only calls out.
- 02 / ConnectConnect the systems your agents change, starting with Stripe. Sigilla reads what really happened there.
- 03 / AskAsk questions in the console, or save them to run on every episode.
- 04 / ProveExport a sealed evidence package when someone asks.
Trust
Built so you don’t have to trust us.
Append-only
Records are never edited. A correction is a new entry that points at the old one.
Encrypted before it leaves you
The collector encrypts in your environment and only makes outbound calls.
No model decides
Findings come from deterministic code. AI can help phrase a question, never answer it.
Open-source verifier
Anyone can read the code that checks your evidence and run it themselves.
Nothing acknowledged is lost
Every record is on disk before your systems are told it was received.
Erasure without breaking proof
Personal data can be erased by destroying its key, while the chain still verifies.
Swipe for more →
Start with evidence. Grow into the whole picture.
Observability and governance for agents are only as good as the record underneath them. We’re building that record first.
- NowEvidence for refund agentsBuilding now with our first design partners: OpenTelemetry agents, Stripe, the console and evidence packages
- NextMore systems and approvalsMore systems of record, identity providers and human approvals
- ThenAgent observabilityLive traces, replay and debugging on a record you can trust
- LaterGovernanceControls over what agents may do, checked continuously
Questions
Something else? Write to admin@sigillalabs.com.
Is Sigilla an observability tool?
Not yet. Observability shows you what your agents are doing; Sigilla proves what they did, and what they didn’t. We’re building observability on top of the same record next.
Do I need to change my agents’ code?
Agents that emit OpenTelemetry work as they are. A small tool wrapper adds a correlation id to outbound calls, so each action joins up with the system it changed.
Where is my data kept?
The collector encrypts records inside your environment before they leave it, and only makes outbound calls. We store encrypted records, and personal data can be erased by destroying its key without breaking the proof.
Which systems do you support?
Stripe first. More systems of record, identity providers and human approvals are next. Tell us what your agents touch.
Will auditors and counsel accept the evidence?
Every package is verification-ready: anyone can check it offline with the open-source verifier. Whether it is accepted is up to the auditor or the court.
Does AI decide the answers?
No. A model may help you phrase a question or summarise a result, but deterministic code decides every YES, NO and CANNOT BE ESTABLISHED.
Put your agents on the record.
We’re onboarding a small group of design partners running agents in production. Join the waitlist and a founder will reach out personally.