Steering and perception

Two collaborators, human and AI, work on a shared plan but cannot see each other’s screens. The human sees the sidebar; the agent sees command responses. Steering and perception are how each side keeps the other in view.

Steering is the guidance an exo command returns alongside its result: what just happened, what to do next, and who decides. Under the SOAR loop (RFC 00224) the human decides and the agent acts, so steering never auto-executes. It surfaces ranked options and waits.

Perception is how user feedback, system observations, and plan changes reach the agent at the moment it can act on them, rather than only when it explicitly asks for status.

Every command is a perception touchpoint

The agent spends most of its time executing, not calling exo status. If feedback only arrived through status, the agent would be blind to it during the work where it matters most. RFC 10182 closes that gap: commands that operate on an entity return steering scoped to that entity’s tree (task, then parent goal, then phase).

Consider an agent reporting progress while the user has just pushed back:

exo task log auth-task --log "Implemented OAuth flow"

Task 'auth-task' logged.

⚠ User concern on goal 'auth-system' (12 min ago):
  "I don't think we need OAuth here — simple tokens might be enough"
  → exo inbox list --entity-type goal --entity-id auth-system

→ Next: exo task list (check remaining tasks)

The agent did not ask for feedback. It reported progress, and the system told it what it needed to know before it kept building the wrong thing.

The inbox as a steering channel

The inbox is the channel through which perception events flow (RFC 10181). Every event carries an entity scope (what it is about), a source (user-feedback, system-observation, or plan-mutation), a priority (immediate, next-touch, when-relevant), and an intent (claim, concern, inquiry, or fyi).

Steering delivers signals, not payloads. A command response carries a one-line summary, such as “2 items on goal ‘Widget Refactor’”, and the agent pulls the full detail when ready. The one exception is entity closure: attempting to complete a goal or task with unresolved feedback surfaces that feedback in full, forcing it to be addressed before the work is marked done.

Plan mutations emit perception events automatically. Adding a goal, resolving an inbox item, or promoting an RFC writes the state change and the perception event in one operation, so a change made on one side becomes visible on the other.