What the approval gate is for
最后更新：
Last updated: 20 September 2026
The practical difference between an agent and a chatbot is that an agent's words become real events. The approval gate sits in between: consequential actions stop and ask before they happen.
Which actions get held
The criterion isn't difficulty. It's whether the action can be undone.
| Action | Hold it? | Why |
|---|---|---|
| Reading files, searching | No | No consequences |
| Editing uncommitted local files | Depends | Recoverable if it's under version control |
| Deleting files, clearing directories | Yes | Gone is gone |
| Committing, pushing | Yes | Published is hard to retract |
| Calling external APIs | Yes | The other side already received it |
| Anything that spends money | Always | Irreversible and directly costly |
Holding has a cost
Every confirmation interrupts you. Hold too often and one of two things happens: you start approving without reading, or you turn the gate off. Both mean you no longer have one.
So the goal isn't holding more — it's holding accurately.
Configuring it so it doesn't nag
Hold only what can't be reversed
If you can treat "edit a file" and "delete a file" differently, let edits through and hold deletions. In most work, edits are nine out of ten actions and deletions less than one — holding only the latter takes the noise down sharply.
Work in a clean tree
If every change is under version control and recoverable, "edit a local file" no longer needs a gate. The danger comes from irreversibility, not from the action itself.
Batch the dangerous ones
Rather than approving each deletion separately, have it list what it intends to remove and approve the batch at once. You're reviewing intent, not individual keystrokes.
When to loosen it
Once you know an agent's temperament in a project — you've seen it read before writing, and it hasn't gone off on its own — you can relax the gate for that project.
Base that on observation, not trust. Watch it run a dozen tasks, each time reading first, and you have grounds. Turning it off because "it seems smart" is a different thing entirely.
In Aiglade
Aiglade gathers pending approvals from every agent into one place, rather than making you discover them across four terminal windows. The more agents run in parallel, the more this matters — three simultaneous confirmation requests are far quicker to handle together than separately.
Related: splitting work across agents.
Key numbers
- Approval mechanism: risky operations — deleting files, changing important settings — pause and wait for your confirmation before running, instead of executing directly.
- 4 built-in AIs: Claude Code, Codex CLI, Hermes, OpenClaw.
- Local database: 24 tables, 131 HTTP endpoints.
- Concurrency: up to 10 tasks by default (max 5 per agent), 8 built-in agent nodes.
- Get your gateway key in about a minute after signing up.
- Task and confirmation records stay on your machine; only the actual AI call uses the network.
FAQ
What does the approval gate actually hold?
Not "the hard things" but the irreversible ones: deleting files, clearing directories, commits/pushes, external API calls, anything that spends money. Reading files and searching are never held. Edits to uncommitted local files are "depends" — recoverable under version control, usually not held.
Why are deletions always held?
Because "gone is gone." Edits are nine out of ten actions, deletions less than one — and deletion is the most common irreversible kind.
What if confirmations get too noisy?
Two bad outcomes: you start approving without reading, or you turn the gate off entirely — both equal no gate. Tune for accuracy, not coverage: hold only what can't be reversed and the noise drops.
Several agents asking for approval at once?
Aiglade gathers every agent's pending items in one place instead of across four terminal windows — the more parallel agents, the more this matters.
Call it in 3 steps
- Set the scope: hold deletions, commits/pushes, external API calls, spending; let reads and searches through.
- Work in a clean tree: with everything under version control, local edits no longer need a gate.
- Batch the dangerous ones: list what will be deleted, review once, approve once — review intent, not keystrokes.
- Loosen only after observing behavior: a dozen tasks of "read before writing" first — never turn it off because "it seems smart."