Team step attempts (retry on another model)

  • Desktop app
  • All plans

What it does

A team step can carry an ordered list of attempts, each with its own provider, model and effort, for example Sonnet medium, then Sonnet high twice, then Opus medium. The step starts on the first one and only reaches the next after a failure you chose to count, so the expensive model only runs when the cheap one failed. After the last attempt the run stops and the agent you named gets a report. A step without a list keeps its one fixed model, exactly as before.

Where to find it

Teams window > select a team > click an agent step > properties panel > Attempts (optional), below Check command. On the canvas, a step with a list shows a looping-arrows badge. Each attempt then shows on the run's timeline, with its number and its model, and on the step's row in the run box (for example 2/4 sonnet).

How to use it

  1. Tick Retry on another model when this step fails. The first row starts as the step's own model.
  2. Set each row: provider, model, effort, and how many times in a row it runs. Add an attempt copies the last row, the arrows reorder them. A field left on Step's model keeps the step's own value.
  3. Under Counts as a failure, tick what moves to the next attempt:
    • The check command fails: the step's own check, or a later step's check that sends the run back to this step. A check that was cut off counts too.
    • The outcome is rework: the needsRework flag is true or qaPassed is false, on this step's report or on the report of a later step that sends the run back here.
    • The step crashes: its console is gone, never took its brief or went silent, and stayed that way for 2 minutes. A console that repairs itself in that time costs nothing.
    • The step runs too long: set the minutes per attempt. The console is restarted on the next attempt.
  4. Under After the last attempt, send a report to, pick an agent of the project, or Nobody. The run stops as blocked either way.

The next attempt keeps the conversation: it restarts the console when the model or the effort changes, and the agent is told which attempt it is on and what failed. A first visit to a step always starts on attempt 1, whatever the previous step reported.

Settings

None global. One field on the step, retryPolicy: { attempts: [{ provider?, model?, reasoningEffort?, times? }], triggers: ["check-failed" | "rework" | "crash" | "timeout"], timeoutMinutes?, reportToAgentId? }. Off by default. times goes from 1 to 5, a list holds 12 attempts at most with repeats. New lists tick the check, rework and crash failures; the timeout needs a duration first.

Agent tools (MCP)

  • teams_save: retryPolicy on an agent node; nodePatches: [{ id, retryPolicy: {...} }] adds one to an existing step, null removes it. No attempt, no trigger, or a timeout without timeoutMinutes is refused. A list longer than maxCycles draws a warning.
  • capabilities_get({topic: "teams"}): the stepAttempts section describes exactly what the runner does.
  • team_read_timeline: lists step-attempt and step-attempts-exhausted events. backlog_get reports a stopped run as blocked, its reason names the step and the failure.
  • The field is frozen while a run of the team is alive, like the check command.

Providers

All providers: each attempt can run on a different one, through the same launch path as a step's own model (quota rules and hidden models still apply). Effort only shows for a provider with that lever. A step set to follow the project's agent keeps the fields its attempt pins and takes the rest from that agent.

Mobile

Nothing to set on the phone: team steps are edited on the desktop. The run box on the phone shows the attempt a step reached, and a run stopped on its last attempt shows its reason on the run card.

Limits

  • A ticked failure goes back to the same step before any connection is followed: a connection on the same flag does not fire while that failure is ticked.
  • A retry of the step itself costs no cycle. A later step it loops through still counts its own, so with a list longer than the team's maximum cycles the run can stop on "Max cycles" first.
  • A question to you (a question card, the needsInput flag, a Wait for me step), a human sending the run back and a manual re-entry are never failures: the step keeps the attempt it reached. A question card also pauses the time limit.
  • A retried branch of a parallel split runs without its entry check, which passed when the split started.
  • Stopping a step's console by hand counts as a crash when that failure is ticked: cancel the run to stop it for good.
  • Since 2026-11-04.

Common questions

  • Can a team step use a cheap model first and a better one only if it fails? Yes: give the step an Attempts list and tick what counts as a failure.
  • Does a failed step retry on the same model? Only if your list says so. Without a list, a loop back runs the step's one model again; with one, each failure moves to the next row.
  • Does my question to the owner waste an attempt? No. A question is never a failure and does not burn the time limit.
  • What happens when every attempt failed? The run stops as blocked, its timeline shows every attempt, and the agent you named gets a report. Re-enter the step to run the last attempt again, or cancel the run.