Skip to content
Hendl
PLATFORM  /  HUMAN-IN-THE-LOOP/ people keep the final call

Agents do the work. People keep the call.

Every consequential action stops at a gate you set. Reviewers see the agent's plan, evidence and provenance — approve in one click, or correct it and let the work continue. Nothing happens silently.

1
Approval queue
across every agent
100%
Reasoning shown
plan, evidence, provenance
1-click
Override
correct and continue
0
Silent actions
everything on the record
ONE QUEUE

Every decision waiting on a human, in one place.

Approvals from every agent land in a single worked queue, routed by your rules — case owner, team, or a named role for specific decisions.

Approvals · Your queue3 waiting · oldest 22 min
Set initial reserve at £4,200Low risk
Reserve agent · CLM-4821 · Motorcomparable repairs, no injury
ApproveAdjustReject
Offer settlement of £11,350Referral
Settlement agent · CLM-4796 · Propertyabove your £10k threshold
ApproveAdjustReject
Escalate to stage 2Escalated
Triage agent · CMP-1174 · Complaintvulnerability signal detected
ApproveAdjustReject
Approvals carry SLAs — the agent chases, then escalates. Nothing times out into action.
BEHIND EVERY REQUEST
The plan
What it intended to do, and why
The evidence
Documents and data behind each step
The provenance
Where every figure came from
The alternatives
What it considered and rejected
Reviewers judge the reasoning, not just the answer.

Four ways a human stays in the loop.

Supervision isn't one checkpoint — it's gates before the action, thresholds around it, queues beside it and sampling after it.

Approval gates

Named steps where an agent must stop and hand the decision to a person — set per task and per level.

Referral thresholds

Value, exposure and vulnerability triggers that pull a case out of the flow and route it to the right reviewer.

Exception queues

Anything an agent can't resolve confidently lands in a worked queue with context attached — never a dead end.

Quality sampling

A standing sample of autonomous work is re-reviewed by humans, so drift is caught before it matters.

OVERRIDES

Corrections aren't friction. They're training.

When a reviewer changes an agent's work, the case moves on immediately — and the correction becomes evidence for a better playbook.

Correct in place

Change the figure, redraft the letter, reroute the case — the agent continues with your correction, immediately.

Reason captured

A one-line reason on each override becomes structured feedback for the playbook owner — not a lost Slack message.

Feeds learning

Corrections roll into evaluation runs, so the same mistake gets rarer over time — that's continuous learning.

Questions teams ask about supervision.

Does every action need approval?

No — that's the autonomy dial. Routine steps can run at notify or autonomous; consequential ones stay gated. You choose per task.

Who gets asked?

Routing follows your rules: case owner, team queue, or a named role for specific decisions — with cover for absence and out-of-hours.

What if nobody responds?

Approvals carry SLAs. The agent chases, then escalates up the route you define. Nothing silently times out into action.

Can a reviewer see why the agent decided?

Always. The full plan, evidence and provenance sit behind every request — reviewers judge the reasoning, not just the answer.

Scale the work. Keep the judgement.

© 2026 Hendl Labs Ltd. All rights reserved.
Services AgreementPrivacy PolicyTrust Center