Keeping a human in the loop: where to put sign-offs
A practical way to decide which agent actions need a person's approval, which need a spot-check, and which can run on their own.
Updated 2026-11-19. "(source N)" refers to the numbered list at the end.
Why sign-offs matter
An agent that asks permission for everything is slow and annoying. An agent that asks for nothing will one day refund the wrong customer. The skill is putting sign-offs in the few places where they count.
OpenAI's guide names two moments that should always bring a person in. One is when the agent keeps failing, for example after several tries to understand a customer. The other is any action that is sensitive, hard to undo or high stakes, such as cancelling orders, issuing big refunds or sending payments (source 1). Anthropic adds that agents should pause at set checkpoints or when they hit a blocker, and should have a hard stop such as a maximum number of steps (source 8).
Sort actions by risk
List every tool your agent can use. Then sort each one into a bucket by asking two questions: can it be undone, and who sees the result?
- Runs freely: reading data, searching, drafting text that stays inside your systems. Example: looking up an order.
- Spot-check: low-risk changes that are easy to reverse. Example: tagging a ticket or updating an internal note. Review a sample each week.
- Approve first: anything a customer sees or that changes records others rely on. Example: sending an email reply or editing a price.
- Always a person: money, deletions, contracts, and anything legal or medical. Example: refunds over a set amount, or closing an account.
Use thresholds, not blanket rules
The same tool can sit in two buckets. A refund under $20 might run freely, while one over $200 needs a manager. Put the threshold in your code, not just in the prompt. A prompt can be argued with; a code check cannot.
Start strict and loosen over time. OpenAI notes that human oversight matters most early on, until you trust the agent's track record (source 1). Keep a log of every approval and rejection. If a person approves the same kind of action 200 times in a row, consider moving it down a bucket.
How the tools support this
Most agent frameworks now have approval built in, so you rarely need to build it from scratch.
- OpenAI Agents SDK: mark a tool as needing approval and the run pauses with a list of pending calls. A person approves or rejects each, and the run resumes. The paused state can be saved and picked up later (source 2).
- LangGraph: an interrupt pauses the graph and saves its state. A person can approve, reject or edit, and the graph resumes from the same point (source 3).
- Claude Code: in its Manual mode it asks before editing files or running most commands, and you can set allow, ask and deny rules. Deny rules are checked first (source 4). An auto mode instead has a separate classifier review actions before they run (source 5).
- MCP: the spec says apps must get the user's consent before running any tool, and servers can ask the user to confirm through elicitation (source 6).
Avoid approval fatigue
If people see fifty prompts a day, they stop reading them and click yes. That is worse than no prompt at all. Claude Code, for example, supports allow-lists for safe, common commands for this reason (source 5).
Good approval screens help. Show exactly what will happen, in plain words: Send this email to [email protected] with this text. Refund $84.00 to order 1182. One click to approve, one to reject, and a box to edit.
Review code like code
For coding agents, the pull request is the sign-off. GitHub lets you use branch protection rules to stop its cloud agent from creating or updating pull requests on protected branches, and tells users to review and test the output before it reaches production (source 9). Require a human review on every agent pull request, at least to start.
Watch the reviewers too
Human review is also how most teams measure agent quality. In LangChain's survey, 59.8% of teams use human review in their evaluations (source 7). Use those reviews twice: once to approve the action, and again as test cases for the next version.
If you sell into regulated areas
The EU AI Act requires human oversight for high-risk AI systems. People must be able to understand the system's limits, watch for over-reliance, override or reverse its output, and halt it, for example with a stop button (source 10). The page we cite gives 2 December 2027 as the start date for the high-risk uses listed in Annex III, and 2 August 2028 for those under Annex I (source 10). Even if you are not in scope, it is a good checklist.
A one-page policy
Write down, for each agent: what it can do on its own, what needs approval and from whom, the money or volume limits, and how to switch it off. Review it each quarter. That single page will prevent most of the mistakes that make headlines.
Sources
Written by Agents Lookup in our own words. Facts come from these pages:
- 1. OpenAI: A practical guide to building agents: cdn.openai.com/business-guides-and-resources/a-practical-guide-to-building-agents.pdf
- 2. OpenAI Agents SDK: Human-in-the-loop: openai.github.io/openai-agents-python/human_in_the_loop
- 3. LangChain Docs: LangGraph interrupts: docs.langchain.com/oss/python/langgraph/interrupts
- 4. Claude Code Docs: Configure permissions: code.claude.com/docs/en/permissions
- 5. Claude Code Docs: Security: code.claude.com/docs/en/security
- 6. Model Context Protocol: Specification: modelcontextprotocol.io/specification/latest
- 7. LangChain: State of Agent Engineering: www.langchain.com/state-of-agent-engineering
- 8. Anthropic: Building effective agents: www.anthropic.com/engineering/building-effective-agents
- 9. GitHub Docs: About Copilot cloud agent: docs.github.com/en/copilot/concepts/agents/coding-agent/about-coding-agent
- 10. EU Artificial Intelligence Act: Article 14, Human oversight: artificialintelligenceact.eu/article/14