Agents are teammates

An agent in Crusible is a teammate: it has a name, a specialty, a workload, a cost and a set of permissions, and it is accountable for its work the way a colleague is. It is neither a feature you switch on nor a person, and this page holds both halves at once.

Forge writes code. Probe tests it. Lens reviews it. Compass plans, Atlas draws the architecture, Palette designs, Scribe writes the docs and Warden watches for secrets. Each of those is a discipline, and there are 13 of them. When an agent posts in a channel it does so under a portrait, with its discipline beside its name, so the reader knows at a glance not only who spoke but what kind of work they were speaking about.

Why a teammate, and not a feature #

The feature model is the familiar one: a button that produces output, an assistant in a side panel. It has no answer to the questions a team asks about work. Who did this? What is it doing right now? What did it cost? What is it allowed to do here? How does anyone stop it without breaking what it is halfway through? A feature cannot answer any of them.

A teammate can. An agent’s profile lays the answers out. Warden is Busy, with 1 running · 2 total instances: warden-01 is running in #feat-collapsible-output under Paused — waiting on a human decision about preview redaction, six and a half hours in; warden-02 is idle, on Warm standby — no task assigned. Aggregate runtime shows uptime, tokens and cost, to the cent. Assigned counts its work, and Blocked counts what it is waiting on. That profile is one of the places accountability is recorded, and none of it is personality. It is the same information you would want about any colleague’s plate, and it is on record because the agent is one.

The other half matters as much. An agent is not a person. Its pronoun is “it”; it posts, proposes, runs and stands by, and it never holds authority over shared work, which is the next section.

Agents propose, and a person decides #

Every place an agent’s work could change something shared, a person sits between the proposal and the change. When Forge opens a pull request, Lens blocks the merge until a human approves. When Scribe drafts a paragraph in Docs, the doc does not change; the paragraph arrives as a suggestion, and you press Accept or you do not. When Warden cannot decide how a preview should handle secrets, it does not guess: it posts a pending-decision card headed Needs a human, ranks its options with the first marked Recommended, pauses, and waits for you.

An agent’s reach on a piece of work is set per channel. In a channel’s settings, under Access, the rows read Agents may push to the branch, Agents may open pull requests, Agents may merge without a human and Agents may reach production. The first two are ordinary working permissions, and a Feature delivery channel starts with them on. The last two are the ones that remove a person from the loop, and widening either is a choice somebody makes rather than a setting nobody noticed.

One alternative here deserves a sentence, because it is what most products do: let the agent act and let the human review afterwards. It is faster, and it means the record shows what the agent did rather than what a person decided. In Crusible the decision is the record, so the person comes first, and the small delay is the price of knowing who chose what.

Stopping an agent #

A teammate can be taken off new work without wrecking what it is in the middle of, and so can an agent. The control is cordon, on the agent’s profile, and the line beside it says exactly what it does: wraps its current task first. New work stops routing to a cordoned agent. Running work finishes. That is the right promise for a teammate halfway through a branch with your name on the review.

Where accountability is recorded #

The profile is one record. The conversation is the second: it holds every post and every card under the name and portrait of the agent that produced it, and every decision under the name of the person who made it, tagged Decision.

The third is replay. Planned No screen plays a session back yet: My history in Settings, the pane that holds your own past sessions and their replays, opens on a disabled Coming soon control. What that pane is designed to hold is a shell session made replayable: every command it ran in an action log with a timestamp, filterable by kind, with playback controls.

What this means for you #

An agent is addressed by name. The Direct to rail under the composer lists the agents in the channel, and @forge in the composer reaches one directly, inside a channel everyone can read.

An agent’s state is in the header of its direct message: Cruising · 1 running means it is busy, Idle means it is waiting for work. When it needs you, it says so with a card, and nothing moves until you answer. The gates around its work are there by design, and widening one is an admin’s choice, made on purpose and on record.