What is delegated approval for AI agents?
Delegated approval is a pattern in which an AI agent pauses at a decision that needs a person, saves its state and releases compute, then resumes from that state once the approver responds, even days later.
Checked
An approval that must happen within the same session, within seconds, pushes teams towards rubber-stamping or towards removing the gate. Real approvals rarely work that way: a finance lead signs off tomorrow, a manager reviews after a meeting. Delegated approval separates the agent's run from the approver's response time.
The term comes from Google's May 2026 guidance on production agents, which describes workflows where the agent pauses for human review while consuming zero compute, alongside Agent Runtime support for agents that keep state for up to seven days. The same mechanism exists, under other names, in most agent SDKs.
What has to be true for it to work
State outside the model
The agent's current step lives in durable storage, not in chat history. Google's ADK onboarding example keeps an explicit state machine in a persistent session store, so a restarted container resumes from the last checkpoint.
A resumable run
Pending tool calls can be saved and restored in another process. The OpenAI Agents SDK serialises its run state to JSON for exactly this.
A wake-up signal
A webhook or queue message resumes the run. In the ADK example, webhooks such as a document-signed event apply the state change before the agent's next step.
Idempotent actions
OpenAI's SDK documentation warns that approval does not guarantee exactly-once execution: a crash after a tool acts but before state is saved can repeat the action on resume. Stable idempotency keys prevent the duplicate.
How SDKs expose it
In the OpenAI Agents SDK, a tool marked needs_approval pauses the run and returns a list of interruptions. The application approves or rejects each one on the saved state and resumes the run, and a decision can be made sticky for later calls to the same tool.
Google's May 2026 ADK post on long-running agents treats waiting on a person, such as a signature on onboarding documents, as a pause in an explicit state machine, resumed by a webhook. It also names the failure the pattern prevents: an agent that replays a long history after an idle gap may assume approvals that were never given.
Securing the approval itself
Trusted state only
OpenAI's documentation notes that serialised run state is not authenticated, so only trusted snapshots should be restored, and the full snapshot should stay on the server.
Check the reviewer
Authenticate and authorise whoever makes the decision, not only whoever receives the notification.
Block replays
Validate each decision against the stored pending request and consume it atomically, so one approval cannot be used twice.
Show consequences
The approver should see the refund amount or the email text, not a raw tool call, or the gate turns into approval fatigue.
How Motiontic's engineers apply it
Motiontic's engineers build approval gates as durable pauses from the start: state in the client's database, approvals routed to the named owner through tools they already use, such as email or chat, and resumption keyed to the decision. Every approved action carries an idempotency key, and approvals and rejections reach the audit log with the approver's name.
In-session approval compared with delegated approval
| In-session approval prompt | Delegated approval | |
|---|---|---|
| Who approves | Whoever is watching the session | The named owner of the decision |
| Time to respond | Seconds to minutes | Hours to days |
| Compute while waiting | Session held open | None; state saved and run released |
| Main risk | Approval fatigue and rubber-stamping | Replayed approvals and duplicate actions on resume |
| Fits | Interactive coding and desktop agents | Back-office workflows such as refunds, onboarding and procurement |
Common questions
What is delegated approval?
Delegated approval is when an AI agent pauses for a human decision, saves its state and stops using compute, then resumes from the saved state once the person approves or rejects.
How does an agent pause without using compute?
Its state, including the pending tool call, is written to durable storage and the process ends. A webhook or queue message carrying the decision starts a new process that loads the state and continues.
What happens if the agent crashes after an approval?
The action may run twice on resume unless it is idempotent. OpenAI's Agents SDK documentation recommends stable idempotency keys or checking the destination system before retrying.
Is delegated approval only a Google feature?
The name comes from Google's guidance, but the pattern is available elsewhere, for example through approval interruptions and serialisable run state in the OpenAI Agents SDK, and through any framework with durable execution.
Related
Human in the loop
Approval gates where a person decides, placed by risk rather than on every step.
Approval fatigue
Too many approval prompts, and people stop reading them.
Long-running agent
Agents that work for hours or days, and the state that lets them pick up where they stopped.
Durable execution
Runs that survive crashes and pauses because every finished step is recorded.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.