Free · no account · every claim answerable by the party it names
What is this agent's risk?
Before you ship on a framework, a model, or somebody else's agent — see what has actually gone wrong with it. Free, no account, and every claim is answerable by the party it names.
Nothing has been reported yet. This registry publishes what strangers file and what adjudication confirms — it prints no number it has not earned, so there is no number here. File the first one, or read what we will not do with it.
Built on
Every tier prints the rules that fired and the numbers they read, so you can recompute it from the page rather than trust it — and every export is signed over its canonical form, so you can check a track record offline without trusting agitome to tell you the truth about it.
What agitome will not do
A report is a signal. It is not a finding, and we will not print it as one.
Most of this registry is what strangers have reported. Some has been adjudicated against a signed loss event. Those are different claims about the world, and every page says which one you are looking at.
Filed by the public. Published, searchable, and it moves no agent's score.
Adjudicated, with a signed loss event behind it that agitome resolved and confirmed.
No rankings.
Two models half a point apart, over populations of different sizes, are not first and second at anything.
No incident rate.
No denominator exists for how many deployments there are, so there is no rate to publish.
No prices.
The calculator does arithmetic on your numbers. It is not a quote, and will not be until the loss data exists.
What you can do with it
Twelve things, all free, no account.
01
Search the incident database
Every reported agent failure, by framework, model, category, vendor or DID. Each has its own page and its own link.
02
Read a risk profile
“Is LangChain risky to deploy?” — the counts, the failure modes, the tier, and every rule that produced it.
03
Pull a risk report
One report across your whole stack. Shareable, and a snapshot: the link shows what you saw.
04
Run the admission gate
Your policy over reported data. Allow, review or deny — before the thing reaches production.
05
Estimate your exposure
What an agent could cost you if it fails, and an honest answer on whether anybody could underwrite it yet.
06
File an incident
Two fields, under a minute, no account. It enters as an unverified signal and is queued for adjudication.
07
Browse by failure mode
The closed taxonomy every product in the portfolio classifies against — so counts are comparable across subjects.
08
Read the State of Agent Risk
The periodic drop off the whole corpus, with the methodology printed beside it so it can be quoted in full.
09
Verify a track record
Check a signed track-record export offline — without trusting agitome to tell you the truth about it.
10
Dispute or claim a subject
This page is about your product. Four things you can do, and three of them need no account at all.
11
Embed the risk badge
A flat, cacheable badge for your own directory or docs. An image — no runtime of ours on your page.
12
Read about attestations
ERC-8004 attestations, for agents whose standing has to be readable by a contract rather than a person.
If this is about your product
Anyone can dispute an incident. You can do more than that.
Claim the subject, and your dispute takes an incident out of every published statistic while contested — it stays visible, because hiding a contested claim is what makes a registry unauditable. If we decide against you, you can still attach a permanent statement.
And if we miss a deadline
Disputes are adjudicated in 30 days, 45 if we record an extension. Past that, the dispute resolves in your favour automatically and the incident stops counting.
The party that failed to decide in time is agitome, and the cost of that lands here rather than on you.
The monthly drop
What broke, what got fixed, and what the data cannot yet say.
Tell me when this changes
We record your address so we can tell you when this changes — and a person reads it.