What you can send to a human.

01 — Review a document

task.type = "document.review"

Something a model shouldn’t sign off on alone — a contract clause, a medical note, a flagged edge case. A trained person reads it, applies judgement, and returns a decision with reasoning.

Typical uses

• Contract clause review
• Medical or clinical note check
• Flagged edge-case adjudication
• Quality assurance on model output

02 — Make a phone call

task.type = "voice.call"

Confirm a booking, chase a supplier, verify a detail. The jobs that still need a human voice on the line. The specialist makes the call and returns the outcome and any notes.

Typical uses

• Booking confirmation
• Supplier or vendor follow-up
• Detail or identity verification by phone
• Appointment chasing

03 — Validate ambiguous data

task.type = "data.validate"

Two records that might be the same person. A field that doesn’t parse. The judgement calls models guess at and get wrong. A person resolves the ambiguity and returns a clean, confident answer.

Typical uses

• Entity / duplicate resolution
• Data deduplication and matching
• Unparseable field interpretation
• Edge-case classification

04 — Take a photo, check the world

task.type = "world.capture"

Confirm something physically exists, capture proof, or eyeball a real-world state your agent can’t see from a server, then send back what was found.

Typical uses

• Proof-of-existence capture
• Physical state verification
• On-the-ground photo capture
• Real-world status check

05 — Make the call you can’t automate

task.type = "decision.own"

The decision that needs real accountability — a person who owns the outcome, signs their name to it, and stands behind it. Not a confidence score. A named specialist makes the call and returns it with their reasoning.

Typical uses

• Accountable approval / sign-off
• Ambiguous escalation handling
• High-stakes judgement call
• Human-owned exception

Not sure which fits?

If a step in your agent needs real-world judgement, a human voice, physical confirmation, or someone accountable for the outcome, it belongs in the loop. Read what human in the loop means, see the integration shape, or just send us the moment your agent gets stuck and we will map it to a task type.