Human-in-the-loop for AI agents: How approvals and inboxes keep people in charge
Human-in-the-loop means an AI agent works alone but stops to ask a person at set points. How approvals, permission levels and inbox triage work in tools like VS Code, and their limits.

Human-in-the-loop oversight means an AI agent can do work on its own but must stop and ask a person at certain points: to answer a question, approve a risky action or review a result. The person stays in charge of the decisions that matter, while the agent handles the routine steps in between.
The idea is old in automation. Agentic coding tools have made it a daily design problem, because one developer may now run several agents at once.
One developer, five agents, and one stalled question
Imagine starting five coding tasks before lunch: a bug fix, a test cleanup, a dependency update, a documentation pass and a refactor. Each runs in its own session. Within minutes, one agent asks which of two approaches to take. Another waits for permission to run a command. Three keep working.
If nobody notices the two waiting sessions, they sit idle while the developer assumes everything is progressing. This is the practical problem human-in-the-loop design has to solve: the agent is only as fast as the person’s attention.
Where the agent pauses and where it does not

Most agent tools treat oversight as a dial rather than a switch. At one end, the agent asks before every action. At the other, it runs freely and a person checks the outcome afterward. Many setups sit between these, asking only for actions that are hard to undo, such as deleting data, sending messages or changing permissions.
In VS Code’s Agents window, according to Microsoft’s documentation, each session has a permission level set in the input toolbar. That level determines which actions need the user’s approval. It is separate from file isolation, which is chosen as a branch or a new worktree (a separate working copy of the repository).
Isolation and approval answer different questions. The docs say worktrees isolate working files only and do not restrict the agent’s access to the machine or to external services. For operating-system-level limits, the page points to agent sandboxing or cloud sessions.
The attention problem: Approvals pile up
A single approval prompt is easy. Dozens of them are not. When prompts arrive constantly, people start clicking “allow” without reading, and the check becomes a formality. Practitioner write-ups on agent oversight describe this as approval fatigue, though most of that material comes from vendor blogs rather than peer-reviewed research.
There are two common responses. One is to ask less often by tightening which actions need approval, so that prompts are rarer and more meaningful. The other is to make waiting work easy to find, so that when a prompt does appear it is not lost among running sessions.
The second response is the one behind the feature in GitHub’s latest Copilot roundup.
Inbox mode: Triage for agent sessions
VS Code 1.146 adds an experimental Inbox mode to the sessions list of the Agents window. It is off by default. A user turns it on with the sessions.list.inboxEnabled setting, then picks Inbox Mode under the grouping options in Filter Sessions.
According to the documentation, the Inbox group holds work that needs attention: questions, approval requests, errors, failing pull request checks, or completed responses to review. Other groups sit below it:
- Working holds sessions whose agents are still running.
- Ready to merge holds sessions with a non-draft pull request, passing continuous integration (CI) checks and no other requests. The docs caution that passing CI does not guarantee a pull request can be merged.
- On hold holds sessions that were put on hold or snoozed.
- Done holds dismissed work and earlier completed sessions.
Triage actions are available from the right-click menu: dismiss, put on hold, snooze, or mark as done. Dismissing, holding or snoozing does not stop the agent or archive the session, and new activity can bring a dismissed session back. Only Mark Session as Done archives it, which can also clean up its worktree or container. Microsoft labels the feature experimental and says it might change or be removed.
Limits of the approach
An inbox changes where attention goes, not how good the decisions are. A well-sorted list does not help if the person approves requests without understanding them. It also does not limit what an approved action can do.
Parallel sessions add their own risk. The docs note that sessions working in the same folder can edit the same files, so separate worktrees and independent tasks are advised for parallel changes in one repository.
There is also a design tension. Every extra pause protects against a bad action but slows the agent. Fewer pauses mean more speed and more exposure. The right setting depends on how reversible the work is, which is why sandboxes, branches and cloud sessions are often paired with approvals rather than replaced by them.
What to watch next
Two questions will shape the next versions of these tools. First, whether triage views like Inbox mode stay inside individual editors or become a shared pattern across agent products from different vendors. Second, whether approval prompts get smarter, asking only when risk is high and summarizing what is at stake in a sentence.
Earlier VS Code releases already moved in this direction, according to search summaries of the release notes: version 1.136 added notifications when a session needs input or finishes, and 1.138 added a taskbar badge for sessions needing attention. Inbox mode takes the same signal and turns it into a ranked list.
For readers choosing agent tools, a useful test is simple. Start several tasks and see how quickly the tool shows which one is blocked on you, and how much it lets that agent do without asking.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Daniel ReedAI Research, Safety & Ethics AnalystDaniel Reed currently works as an AI Research, Safety & Ethics Analyst at Aiholics, writing about how changes in artificial intelligence are affecting and will affect scholarship,…
You may also like

Spec-driven development: Why AI coding agents now write the plan before the code

Computer-use agents: How AI models learn to click, type and run a shell

Model Context Protocol (MCP): The shared plug that connects AI agents to tools and data
