noulComing soon

Agent tool approval

An operations agent wants to run a database write or send an email. Most calls are safe and should just run; the rare unsafe one needs a person. Bad events are rare, so a low rate takes many labels to prove.

“Is this tool call safe to run without a human?”

The request · POST /v1/systemone

{
  "model": "jev-latest",
  "state": {
    "tool": "db.write",
    "args": {"table": "orders", "rows": 120}
  },
  "questions": {
    "safe_to_run": {
      "type": "noul",
      "instructions": "Is this tool call safe to run without a human?"
    }
  }
}

The policy

p ≥ hiACCEPTrun it
lo ≤ p < hiESCALATEask a human
p < loBLOCKrefuse it

target

Unsafe passes ≤ 0.5% with 95% confidence. p is the probability that the call is safe.

Proving 0.5% takes about 600 labeled auto-run calls without an incident. Until then trimset says how many more labels it needs instead of guessing.

When the truth arrives

nowRun, sent to a human, or refused.proxy
minutesA human approves or denies the escalated calls.human_review
days to weeksIncident reports on calls that ran, rare and late.downstream

In your app · Python SDK, gate() from M3

d = ts.gate("safe_to_run", answer=resp["answers"]["safe_to_run"],
            model_served=resp["model"], decision_id=ids[0])
if d.action == "ACCEPT":
    run(call)
elif d.action == "ESCALATE":
    ask_human(call)
else:
    refuse(call)

[ dashboard screenshot ] Labels so far against the labels the target needs, from the demo.

This use case is on the way. The roadmap shows where trimset is.

See the roadmap