How is this different from sending a Slack message?

Slack notifies; AXME durably waits, times out, escalates, and ties the human action back to the intent lifecycle and audit.

Can multiple humans approve in sequence?

Yes — chain waits across roles or organizations on one intent or coordinated intents.

Does the agent process stay alive while waiting?

No — the agent can exit; Cloud resumes when the human completes the task.

AXME CLOUD

Add human approvals to any AI agent in 3 lines — not 200

Every approval gate requires: webhook endpoint + email service + polling loop + timeout handler + retry logic + state persistence + audit logging. That's 200+ lines per gate.

AXME provides 8 native HITL task types: approval, review, confirmation, assignment, form, clarification, manual action, and override.

Start free Read docs

Production agents touch money, customers, and compliance. A model should not send the board pack or refund a payment without a human in the path — but bespoke approval stacks do not scale to eight different interaction patterns.

AXME Cloud provides eight native human task types: approval, review, confirmation, assignment, form, clarification, manual action, and override. The agent enters WAITING_FOR_HUMAN; the runtime holds state until the person acts — no polling loop in your code.

Reminders, escalation to a backup approver, and tamper-evident audit attach to the same intent. Finance, legal, and ops see who approved what, when, and on which version of the artifact.

8 TASK TYPES

Native HITL primitives.

Not a generic task queue bolt-on.

Approval

Yes/no with timeout and escalation.

Review

Structured review before proceed.

Confirmation

Explicit confirm step.

Assignment

Route to the right owner.

Form

Collect structured input in-flow.

Clarification

Agent asks; human answers.

Manual action

Human performs external task.

Override

Emergency takeover of agent run.

200-line approval gate → 4 lines

DIY webhook HITL

def request_approval(reviewer, ctx):
  token = secrets.token_urlsafe(32)
  db.insert("approvals", token=token, ...)
  send_slack(...); scheduler.add_job(remind, ...)
  # + webhook, poll loop, audit

AXME HITL

intent_id = client.send_intent({...})
result = client.wait_for(intent_id)
# reminders, escalation, audit built in

Agents should not block while waiting

The agent process can exit during WAITING_FOR_HUMAN. Cloud holds durable state and wakes the workflow when the human acts — no busy loop burning tokens.

WAITING_FOR_HUMAN

Agent running

→

Waiting for human

→

Resume on response

OPERATIONS

Built for production gates.

Timeout + reminder

Nudge at 24h; escalate at 48h.

Audit trail

Every response logged to intent.

Learn more →

Delivery modes

Email, inbox, SSE — same wait.