Approvals Center
How the Approvals tab works for human review of sensitive agent actions.
The admin app includes a dedicated Approvals tab under Deployments -> Agents for human-in-the-loop review.
Use it when an agent is allowed to plan or propose a sensitive action, but a human still needs to decide whether the action should actually run.
What the Approvals tab does
The Approvals tab is the operator inbox for agent actions that were paused by a human approval gate.
It lets reviewers:
- see all pending and resolved approval reviews in one place
- inspect the exact tool call payload before anything executes
- view the triggering session, turn, and nearby transcript context
- approve or deny the action with an optional reviewer note
- keep an audit trail of who resolved the review and why
Where it lives
In the admin UI, approvals are a top-level tab alongside:
AgentsSessionsApprovals
That means approval handling is no longer buried inside a session detail view. Operators can go straight to the queue and work through pending reviews.
What reviewers see
The page is split into two main areas.
Review queue
The left side shows a list of approval reviews with:
- status
- agent name
- turn number
- a short summary of the action the agent wants to take
- request and expiry timing
- inferred risk badges such as destructive, production, external, billing, or sensitive
Pending reviews are sorted by expiry so urgent actions surface first.
Review detail panel
The right side shows the full context for the selected review, including:
- agent and session identifiers
- turn number and iteration
- requested time, expiry, and default timeout action
- the user request that led to the action
- the assistant response before the run paused
- nearby transcript turns for context
- the exact tool calls and arguments awaiting approval
- decision metadata after resolution
Typical lifecycle
The full flow looks like this:
- A user message reaches the agent.
- The model decides to call one or more tools.
- Everstack checks those tool calls against the configured HITL approval rules.
- If a tool call requires approval, the session moves into
waiting_for_approval. - A review is created and appears in the
Approvalstab. - A human reviewer opens the review and either approves or denies it.
- The runtime resumes with that decision.
What happens after a decision
Approve
If the reviewer approves the action:
- the review is marked resolved
- the session leaves
waiting_for_approval - the blocked tool call is allowed to execute
- the agent continues its turn normally
Deny
If the reviewer denies the action:
- the review is marked denied
- the blocked tool call is not executed
- the runtime injects a denial result back into the conversation
- the model can continue with that denial as context
This gives the agent a chance to explain the denial, ask for a safer alternative, or choose a different action.
Timeout
If nobody resolves the review before the timeout expires, Everstack applies the configured default action.
Depending on configuration, that default is usually either:
- approve on timeout
- deny on timeout
Why this matters
The Approvals Center is useful for actions where autonomy is helpful but fully automatic execution would be risky.
Common examples:
- deployment or production changes
- destructive file or infrastructure operations
- billing or customer-impacting actions
- sending external communications
- anything involving secrets, credentials, or admin access
Good mental model
Think of the Approvals tab as the point where the agent hands control to a human.
The model can still reason, plan, and prepare the exact action. The human stays in charge of whether that action is allowed to cross the line into execution.