Secure · Human Approvals
Available nowLet a person approve the actions that matter
Halo pauses higher-impact agent actions until a reviewer approves or rejects them. The run waits at that step, the reviewer sees what the agent is trying to do, and the decision is recorded as evidence.
- approval states, pending, approved, rejected, and expired
- 4
Why it matters
Autonomy needs a pause control
Agents can send an email, update a record, or start a payment in seconds. Teams need a way to let routine work run and hold the actions with real consequences for a person.
- 01
Blocking an action outright slows work that a person would gladly approve.
- 02
Approvals handled in chat threads and email leave no reliable record.
- 03
Reviewers need the agent, the action, the target, and the risk in front of them before they decide.
How it works
From a held action to a recorded decision
Reviewers see the context they need, and the record stays with the run.
- 1
Define the trigger
Write a policy rule that holds an action, such as an outbound email to an external domain, and assign the capability a risk level.
- 2
The agent attempts it
The gateway evaluates the call. When the rule matches, it creates an approval request and pauses that run step.
- 3
A reviewer decides
The reviewer sees the agent, action, target, and risk, then approves or rejects with an optional note.
- 4
The record stays
An approval resumes the run and a rejection ends it. The request, the decision, and the linked evidence checkpoint are recorded.
Capabilities
Approvals built into agent execution
A policy rule holds the action, a reviewer decides, and the run continues or ends, with a record at each stage.
Policy-triggered holds
A policy rule with a require-approval outcome holds matching actions, such as an email to an external domain, until a reviewer decides.
Context for reviewers
A request shows the agent, the action, the target, a risk label, and links to the run and session, so reviewers decide with context.
Decision notes
Reviewers can add a note when they approve or reject, and the note becomes part of the evidence record.
Pause and resume
The run pauses at the held step and resumes from that step after approval. A rejection ends the run.
Expiry that stays closed
Set an expiry for pending requests. An expired approval does not allow the action, and the run stays blocked until a new run starts.
Risk-based approval policy
Choose an approval policy for each agent, from requiring approval for high-risk actions to logging only, and add triggers for specific events.
Expiry behavior
An unanswered request stays closed
A pending approval that nobody answers does not turn into permission. When a request expires, the action stays blocked and the run waits for a new attempt, and the evidence checkpoint records that the approval expired.
- Pending requests hold the run at that step
- Expired requests keep the action blocked
- Rejected requests end the run
- Decisions land in the evidence trail
FAQ
Questions, answered
Any action that policy can describe, such as a tool call, a destination domain, or an operation with a high capability risk level. You write the rule, and Halo holds matching actions at the gateway until a reviewer responds.
A request shows which agent is acting, the action label, the target, a risk label, and links to the run and session. Reviewers can approve or reject and add a decision note that becomes part of the record.
Each request can have an expiry. When it passes, the approval is marked expired, the action stays blocked, and the run does not continue on its own. A new run is needed if the action must be retried.
They slow only the actions you choose to hold. Routine, low-risk calls can run under allow rules while higher-impact actions wait for a person. Starting a rule in audit mode shows how often it would hold work before you enforce it.
Each approval links to an evidence checkpoint that records the policy, run, session, and risk label. Approval decisions also appear in audit logs, which export as CSV and JSON. Reviewers can filter those records by actor, action type, and date range.
Put a halo around the AI your business runs.
See discovery, runtime policy, approvals, and evidence working together on your own use cases.