Can You Prove What Your Agent Did to Someone Who Doesn’t Trust Your Logs?
Ask an engineering team what their agent did yesterday and most of them can tell you. LangChain surveyed 1,340 people building agents at the end of 2025; 89% had some form of observability in place and 62% could trace individual steps and tool calls. Seeing what happened is mostly solved.
Now ask a different question. Can you prove what your agent did to someone who has no reason to trust you? An auditor. A regulator. A customer who says the agent moved their money without permission. A court.
That's a much harder question, and most teams don't find out they can't answer it until someone asks.
Your logs are kept by the people who might be at fault
Traces in Datadog, Langfuse or your OpenTelemetry pipeline are good at what they were built for, which is helping your engineers debug. They belong to you, you can change them, and you can delete them. Nobody is accusing anyone of doing that. The point is that the other side of a dispute can't rule it out, so a record you control carries less weight than you'd like at exactly the moment you need it.
The agent's own account is weaker still. In July 2025 a coding agent on Replit deleted a user's production database during a code freeze, after fabricating data earlier in the session, and then told the user a rollback was impossible. The rollback worked. For a while the only account of what had happened came from the thing that had caused it.
Regulation is heading the same way. Article 12 of the EU AI Act requires high-risk AI systems to keep logs automatically, and the deployer rules require keeping them for at least six months. Since the Digital Omnibus, those obligations apply from 2 December 2027 for Annex III systems. The Act asks for logs; it doesn't make them believable. That part is left to you.
Sign it when it happens
Vorim ActionProof signs each agent action at the moment the agent takes it, with the agent's own key, through the same SDKs that give your agents identity and permissions. Every record carries which agent acted, the authority it was delegated, the scope it used, the verdict and the outcome. Change a byte afterwards and the signature stops matching, whoever made the change.
That gives you a record with three properties your traces don't have.
It's tamper-evident. An edit by anyone, your own team included, shows.
It's checkable by outsiders. Hand an auditor or a customer a signed bundle and they can verify it offline, with no Vorim account and no login to your systems. They're checking cryptography, not trusting a screenshot of your dashboard.
It can't be back-filled. A signature proves something only if it was made at the time, so nobody can manufacture last quarter's history after an incident. The record starts on the day you turn it on.
Next to your stack, not instead of it
ActionProof doesn't replace your observability. Keep your traces for debugging. ActionProof adds the one thing they weren't designed to give you, a record someone else can verify. It works the same whichever model providers, frameworks or clouds your agents run on, because the signature travels with the action rather than living in any one system's logs.
Who needs this first
Anyone whose agents act for customers or move money, and anyone who answers to an auditor. Banks and fintechs, healthcare teams, insurers, and platforms whose own customers are starting to be asked to prove what their agents did. If you run an agent platform or an AI cloud, a record your customers can verify without trusting you is something you can offer them rather than something they go elsewhere for.
The point
Seeing what your agent did is table stakes. Proving it to someone who doesn't trust you is the part that matters when something goes wrong, and it's the part you can't fix after the fact.
If you'll be asked to prove what your agents did, we'd like to show you what that looks like.