One agent doing one thing is easy to follow. The trouble starts when agents start working together: one hands a task to another, that one narrows the scope and passes it on again, and three hops later something happens that nobody can quite trace back. The audit answers it eventually, offline, after the fact. But "eventually" is a strange place to leave the question of what your agents are doing right now.
So we built two things. One is about watching it happen. The other is about tightening what each agent is allowed to do in the first place.
Live multi-agent sessions
When a chain of agents works together through a hand-off, that chain is a session. Until now you could reconstruct it afterwards as a signed provenance graph, which is exactly what you want for an audit. What you could not do was watch it live.
Now you can subscribe to a session and every action streams to you as it runs: which agent acted, what it did, and whether it was allowed. An agent hands off, the receiver picks up under a narrowed set of permissions, it acts, and you see each step arrive in real time rather than reading it back an hour later.
The important part is what it does not leak. The stream is keyed to the delegation chain and scoped to your organisation, over an authenticated connection. A session belonging to another company is never visible to you, and yours is never visible to them, even if someone guesses a chain id. And the live view is the same session you replay afterwards as a signed record. It is not a separate log that might drift from the real one. It is the real one, arriving sooner.
This is the difference between trusting that a multi-agent workflow behaved and watching it behave.
Capability profiles
The second one comes from a habit I have watched every team fall into. You grant an agent a set of scopes when you build it, usually a bit more than it needs because you are not sure yet, and then you never revisit it. The grant made sense on day one. Ninety days later the agent has never once used half of it, and that unused half is pure blast radius sitting there for no reason.
A capability profile fixes that by looking at evidence instead of guesses. Vorim compares the scopes you granted an agent against the scopes it has actually used in its real audit history, and hands you back three things: the scopes it uses and should keep, the scopes you granted that it has never touched and could revoke, and any action it took that had no matching grant, which is a gap worth a closer look. The suggested set is simply what the agent really does. Least privilege, derived from behaviour rather than from how cautious you were feeling the day you set it up.
import createVorim from '@vorim/sdk';
const vorim = createVorim({ apiKey: process.env.VORIM_API_KEY });
const profile = await vorim.getCapabilityProfile('agid_inbox_triage', {
windowDays: 30,
});
console.log(profile.suggested); // the tightest set that still covers it
console.log(profile.unused); // granted, never used -> candidates to revoke
console.log(profile.usedNotGranted); // used without a grant -> review thisOne deliberate choice: it changes nothing on its own. The analysis is read-only. It tells you what it would tighten and leaves the decision to you, because silently revoking a scope an agent turns out to need at month-end is its own kind of outage. You get the recommendation, with the evidence behind it, and you make the call. It is in the Python SDK too, sync and async.
Same thread, two ends of it
These are not really two separate features. They are the same question at two moments. Live sessions answer "what is this team of agents doing right now, and can I watch it." Capability profiles answer "given what they have actually been doing, what should they even be allowed to do." One is the real-time view, the other is the quiet tightening afterwards. Both come down to the thing we keep building toward: an agent should only ever do what you allowed, and you should be able to see it, prove it, and trim it.
Both are live in the SDK today.
Ready to build with AI agents?
See how Vorim gives your agents identity, permissions, and a signed audit trail. Book a walkthrough with our team.