Amiril learns what each of your AI agents normally does, flags the runs that don't fit, and explains what went wrong. Where a rule would have prevented it, Amiril drafts one. Your agents check it before every tool call, in microseconds, and it deploys within seconds with no changes to your agents.
Watching refunds-agent, 1284 runs today
When an agent gets something wrong, it rarely crashes. It refunds the wrong order or skips an identity check, then reports the run as done.
Your logs show thousands of completed runs. The wrong refund is somewhere in there.
The three runs least like their agent's usual behaviour, most unusual first.
Amiril learns how each agent usually behaves from its own past runs and scores every new run by how different it is. No AI model decides which runs are unusual.
A higher score means the run is rarer for this agent. Connect your traces and scoring starts, with no thresholds to tune.
You can see why every run got its score.
Statistics pick out the few runs worth a look. An agent then explains what happened in each one, so the expensive part only runs where it matters.
An agent drafts the rule from the unusual run using our statistical models and evidence, in plain text you can read.
when calls "issue_refund" unless exists "verify_identity"[status ok]
Rules are settings, not code. Running agents pick them up within seconds, the way they pick up feature flags.
An agent drafts the rule from the unusual run.
See which of your past runs it would have blocked.
It checks live runs and records what it would block, but blocks nothing.
Your running agents follow it within seconds, with no release.
Each tool call is checked in your agent's own process, so there is no proxy and no network call. The same rule works on any cloud and every framework Amiril supports.
Your agents can change real data in production. Watch them as closely as anyone else with that access.
Amiril logs every tool call it checks, whether it was allowed, flagged or blocked, and the rule that decided it.
Each rule is plain text, so anyone can see exactly what it allows and blocks.
Coming soon. Pause a tool call until a person approves it.
Upload the traces you already have and Amiril shows you every unusual run in them, as far back as your history goes.