human in the loop ai agents: approval gates, questions, and a review gate

human in the loop means an agent stops and waits for a person at the points that matter, instead of asking forgiveness afterwards. in aldena that is not a setting buried in one screen: every tool an agent can call carries a policy you set, agents can pause a run to ask you a question in plain language, and engineering work arrives as a pull request that no agent is allowed to merge. you decide how much rope to give, per tool, per room.

three policies, per tool

policies are set per room, in a Permissions tab that groups tools by where they come from: the built-in ones for files, commands, web access, and memory, plus whatever each connected integration brings with it. set a whole group with allow all, ask all, or deny all, then override the individual tools you feel differently about.

when a tool is set to ask, the agent stops and shows the exact call it wants to make. answer allow once, always allow, or deny, and the run continues from exactly where it paused. always allow writes the policy back for you, so a decision you have already made is not put to you twice.

agents ask you questions

asking a person a question is itself a tool, and it is the one tool that can never be switched off. an agent that hits an ambiguity stops and asks rather than guessing: which environment, which repository, is this in scope. the analyst role is built around it and asks its clarifying round before it writes a line of the spec.

a paused run is durable, not a held-open connection. answer an hour later and the agent resumes with everything it had. that is what makes the pause safe to rely on.

the review gate holds

for engineering work the last step is deliberate: the delivery manager branches, commits, pushes, and opens a pull request against your default branch, then stops. no agent merges its own work. the diff, the test results, and the reasoning are all in front of you before anything reaches your main branch.

the same shape shows up elsewhere. the scrum master presents the whole proposed backlog and waits for an explicit go-ahead before it creates a single ticket in your tracker. scheduled work and event subscriptions an agent set up for itself are listed in room settings so you can see what is on the books and cancel it.

the concierge asks too

the built-in assistant that sets up teams, rooms, agents, and billing on your behalf runs under its own permission list in your account: teams and members, rooms, agents, skills, technologies, environment variables, scheduled tasks, room tools, billing and credits, and account settings. same three policies, same prompt before it acts.

pair this with isolated rooms, which bound what any approval could reach in the first place, and with the org chart, which decides who was doing the work when the prompt appeared.

faq