Why 'Outcome unknown' cannot be a retry button
When a side-effecting connector drops after dispatch, retrying blindly duplicates actions and marking it failed hides partial work. Why workflows and swarms need six execution states instead of two.
Most workflow engines collapse execution outcomes into two buckets: green when a response arrives with a 2xx code, red when anything times out or drops. Once a step turns red, the standard reliability feature is an automatic retry loop. For a pure read or a stateless model completion, that works. For any call that changes external state—sending an email, charging a card, posting to an issue tracker, firing a webhook—treating a lost confirmation as “failed, try again” is a bug with a blast radius.
While compiling our 2026 AI Experience and Workflow Research report, visible execution and recoverable work ranked #3 overall across 49 coded developer discussion units (16 supporting units, high confidence in the need). Three distinct failure modes showed up repeatedly across workflow builders and coding agents: blind retries duplicating side-effects [WF08], aborting one stuck action wiping out completed sibling work [UX06], and pending approval prompts locking out steering or accidentally treating a queued follow-up message as consent [UX07].
A timeout after dispatch is not a failure. It is an unknown.
In November 2026, an n8n community report documented a workflow configured with node-level retry-on-fail that re-sent 49 emails after they had already gone out, then marked the overall execution green [WF08].
Once bytes leave your process toward a non-idempotent endpoint and the socket closes before a receipt returns, only two things are true: the request left, and you do not hold proof of whether it landed. Retrying blindly guesses “no.” Marking it green guesses “yes.” Both guesses lie to the operator.
// Read-only or idempotent step: safe to auto-retry with backoff
if (step.effectClass === "read_only" || step.idempotencyKeyConfirmed) {
return retryWithBackoff(step);
}
// External side-effect after dispatch without confirmation: NEVER auto-replay
if (step.dispatchedAt && !step.destinationConfirmed) {
return transitionStep(step.id, {
state: "outcome_unknown",
autoRetry: false,
nextAction: "reconcile_destination_before_replay",
});
}Six states instead of a binary spinner
To keep operators from guessing what happened—or losing half an hour of good output when one step stalls—Section 6 of our specification standardizes six explicit states across chat, workflows, and parallel /swarm runs:
| State | What actually happened | Required controls |
|---|---|---|
| Proposed | A plan or draft exists; nothing external has executed | Edit, inspect, discard |
| Waiting for approval | A specific consequential action is blocked on human review | Approve exact payload, edit, skip, stop (steering stays unlocked) |
| Running | Execution is progressing inside the approved boundary | Inspect live inputs/outputs, stop, redirect |
| Needs attention | A known error prevents useful completion (pre-dispatch or verified error) | Explain cause, repair input/auth, resume eligible work |
| Outcome unknown | Request left the boundary; destination confirmation is missing | Inspect dispatch log, reconcile destination, avoid blind replay |
| Completed and checked | Required deterministic checks and destination confirmations passed | Open deliverable, inspect evidence, reuse |
Three controls that look minor until a 10-minute run breaks at minute 9
- Separate stopping one action from discarding completed work. When a four-step workflow or a four-worker
/swarmrun has already verified three artifacts and the fourth worker hangs on an external page, pressing stop on that worker must checkpoint and keep the first three outputs intact [UX06][LA10]. Tearing down the whole run because one branch stalled punishes the user for intervening. - Keep steering unlocked while approval is pending—without auto-approving. In Anthropic’s observational autonomy analysis across millions of sessions, experienced users both auto-approved more routine sessions (roughly 20% to over 40%) and interrupted active turns more selectively (roughly 5% to 9% of turns across experience groups) [S09]. When an agent pauses at
Waiting for approval, the human often spots a parameter they want adjusted [UX07][WF07][LA05]. Typing a correction must redirect the plan and invalidate the stale payload—never act as an implicit “yes” to the queued command. - Inspecting prior versions must never mutate the active draft. Clicking an earlier workflow revision or run snapshot to see how a prompt was phrased yesterday must open read-only until the user explicitly chooses to restore it [WF14].
None of these states make external APIs more reliable. What they change is what happens in the thirty seconds after an API misbehaves: completed work stays saved, unconfirmed side-effects are quarantined instead of multiplied, and a green badge means checks actually passed. Read the full methodology and all 65 primary records in AI Experience and Workflow Research (2026).