Approval gates that suspend the run
Most approval UX asks a human to review work that already happened. A real gate suspends the run at the boundary and resumes it only on approval.
What is human approval gates for AI agents?
Most approval UX asks a human to review work that already happened. A real gate suspends the run at the boundary and resumes it only on approval.
Implementation
- Model the run as a graph with suspend and resume points at every irreversible action.
- Persist the suspended state so an approval hours later resumes the same run.
- Show the approver the exact payload that will execute, not a summary of it.
- Record approver identity, decision, and timestamp in the audit log.
- Expire pending approvals rather than leaving runs suspended indefinitely.
Failure modes
- Notification-only gates, where the action already ran.
- Summarised payloads, which hide the parameter that mattered.
- Approvals granted by the same service account that requested them.
How it is verified
- Kill the process while a run is suspended and confirm it resumes correctly.
- Confirm the approved payload byte-matches the executed payload.
Questions this pattern answers
What is human approval gates for AI agents?
Most approval UX asks a human to review work that already happened. A real gate suspends the run at the boundary and resumes it only on approval.
How do you implement approval gates that suspend the run?
Model the run as a graph with suspend and resume points at every irreversible action. Persist the suspended state so an approval hours later resumes the same run. Show the approver the exact payload that will execute, not a summary of it. Record approver identity, decision, and timestamp in the audit log. Expire pending approvals rather than leaving runs suspended indefinitely.
How is approval gates that suspend the run verified?
Kill the process while a run is suspended and confirm it resumes correctly. Confirm the approved payload byte-matches the executed payload.
What usually goes wrong with approval gates that suspend the run?
Notification-only gates, where the action already ran. Summarised payloads, which hide the parameter that mattered. Approvals granted by the same service account that requested them.
Projects in this track
ClawKeeper →
Governed Claude agent control plane. A deterministic policy engine runs before every LLM call, autonomous spend defaults to zero, and the audit log is enforced by the database.
TechTide Harness Kit →
Trust contracts for agent skills expressed as JSON Schema: risk tiers, tool permissions, deny-by-default egress, approval gates, and prompt-injection fixtures behind a wall of CI validation gates.
CipherClaw →
Deterministic multi-agent debugger with zero LLM calls in the analysis path: causal graphs, weighted root-cause scoring, and failure prediction.
Swarm 357 →
357 Claude agent roles across six business layers, with portable memory, enforced cost caps, a bash policy gate, and durable checkpoints.
Related patterns
This page documents how the pattern is built. Engagement scope and pricing live on TechTide AI.
AI Production Systems at TechTide AI→© 2026 Alex Cinovoj · TechTide AI · Columbus, OH