Run and monitor

How to run a workflow, watch it execute, pause and resume, stop a runaway, and recover from a crash.

Running a workflow

Two ways:

Click the Run button on the workflow's control bar (also available from the World View rail and per-workflow frame controls).
Click an individual agent's play icon to run just that agent on its own.

Clicking Run first checks the basics — a provider is configured, the workflow is saved, and it has at least one agent — then starts the run immediately with the workflow's saved configuration.

If you'd like to compose input for each run first, turn on the Input chip next to the workflow's name. With it on, Run opens the pre-run launcher: a panel pinned to the left of your workflow where you can type an instruction for the run, attach files, accept suggested context, and pull in connected data sources. Click Send to start the run.

The launcher is non-blocking: you can keep panning the canvas and clicking agents while it's open. Escape or the × closes it, and turning the Input chip off dismisses it too.

The Overview

A run window opens on its Overview: the shortest honest answer to "where is this run?".

At the top, a strip of counts — running, waiting for you, under review, done, sent back, failed, check by hand — and what the run has spent so far (an amount smaller than a cent is shown to four decimals rather than rounded away to zero).

Below that, one tile per agent. A tile carries the agent's name, one word for where it stands (Running, Under review, Done, Sent back, Failed), what it has spent, how many pieces of evidence it put on the board, how many of its questions are still waiting on you, and — if a review was asked for — where that review stands (Waiting for its reviewer, Reviewed, Sent back, Nobody checked this). An agent nobody asked to review says nothing about review at all, rather than something that would read as a judgement.

Click a tile to see just that agent's lines in the log; a Show everything control brings the whole run back.

Under the tiles, Cerevisor noticed collects anything worth saying out loud about the run as a whole: a tool that failed five times in a row, a credential that was switched away from twice, an agent that hit the same error twice running, or the note that Cerevisor closed before this run finished. These are observations. Cerevisor tells you and keeps going — it never stops a run because of one.

Two honest cases:

A run that has not started yet says No agents have started yet rather than showing an empty grid.
A run finished by an older version of Cerevisor says This run finished before Cerevisor kept an Overview. Its log, files and board are still here — its other tabs still hold everything.

Check by hand

Some actions end without an answer. A command was killed by its own timeout after it had started; a web request that changes something was cut off after it went out; a connected server stopped responding mid-call. Cerevisor genuinely does not know whether the thing happened, and from 3.2 it says so instead of guessing.

When that happens:

The agent is told, in plain words, that the call may have gone through and must not be repeated.
Cerevisor makes sure it is not repeated in that turn: the very next identical call is refused before it can run. If the agent is later retried from the start, the count begins clean.
The step is added to the run's check by hand count, and shows in the run window under Cerevisor could not tell whether these ran. Only you can say.

Each row asks one question — Did the [tool] call go through? — with two buttons: It happened and It did not happen. Your answer is written into the run's own history as your decision, so anyone reading the run later sees what you established.

Once you have answered, the row reads You said: It happened (or It did not happen).

You can change your answer. An answered row carries a Change my answer button that offers the two choices again. Pressing one of the plain buttons a second time still does nothing — it says You have already answered this one. To change it, press "Change my answer" — because that is the press a row you have already settled invites by accident. Changing an answer never erases the first one: Cerevisor writes a second decision saying what now stands and what it replaced, so the row reads You said: It did not happen (changed from: It happened) and the run's history shows both. Answering with the answer you already gave is refused as That is already your answer.

On your phone, you can answer but not change. The Companion offers the two buttons on an open row; once you have answered, the buttons are gone and the row just reads You said: …. If you press a button anyway on a row someone else just settled a moment earlier — for example from the run window — the phone tells you where the change can be made instead: "To change an answer, use the run window on your computer."

Cerevisor also refuses to guess in three situations, each with its own sentence rather than a silent dead button:

The run is still finishing — its history is still being written out. Try again in a moment. A run that stopped long ago because the app was closed or killed is closed out first, so it can be answered instead of saying this forever.
The run's history is incomplete — usually because Cerevisor was cut off part-way through writing it. Which step is which would be a guess, so nothing here can be settled. This is permanent for that run, and it applies to exactly the crash-interrupted runs that most invite the question. There is no version of this you can retry.
The step belongs to a different run, or cannot be found in this one.

Once a run starts, the modal shows live status. Top section: overall progress and run cost.

Below that: one card per agent, grouped by wave. Each card shows:

Status badge: Waiting / Running / Completed / Failed / Blocked / Skipped / Stopped.
Streaming output preview (the agent's response as it's generated).
Tool calls in flight, with the tool name and arguments.
Wall-clock time so far.
The agent's resolved provider + model.

Click any agent card to expand it for full output and full tool-call history.

The modal also shows:

Current wave indicator at the top.
Run cost so far (per provider).
A banner when agents are waiting on you — click it to jump to the agent that needs an answer. The actual permission and approval prompts appear on the waiting agent's card on the canvas, not in this modal.

The same statuses show as badges on the agent cards on the canvas. Blocked (amber) means the agent stopped and reported what it needs to continue — a graceful stop, not a crash. Hover the badge to read what it asked for.

Stop, pause, resume

Three buttons in the modal:

Button	Behavior
Stop	Hard stop. All running agents are cancelled; the run ends. Use for true emergencies.
Pause	After each agent's current tool call finishes, the orchestrator pauses. Resume picks up where you left off.
Resume	Continue from a paused state.

The Execution Modal can be closed with Escape: the run continues in the background. To reopen, click the status indicator in the status bar.

Approval modes and permission prompts

When Cerevisor asks for permission is controlled by the approval mode — a three-position switch you'll find in Settings and in the chat builder's "Permission posture" strip (a workflow can also pin its own mode):

Mode	Behavior
Oversight	Every tool call — including safe reads — asks for your approval. Best when you're new to running agents or the workflow touches sensitive files.
Auto (default)	Reads, searches, writes, edits, and safe commands run automatically; consequential actions (deleting files, sending anything, untrusted integrations) ask first.
Hands-free	Never asks. Agents start immediately. Needs a one-time confirmation to switch on; the capability limits in Settings (files, bash, web, git) are still enforced.

When a prompt does appear, it shows up on the asking agent's card. It explains in plain language what will happen, why it's asking (e.g. "Auto mode checks consequential actions before they run"), and shows the workflow's policy at a glance. Options:

Button	Scope
Approve once	This single call.
Always allow [tool]	Every call of this tool for the rest of this run.
Approve all N pending	When several agents are waiting on the same kind of action, approve the whole batch at once.
Deny	Returns an error to the agent.

Permission prompts pause only the agent that asked, other agents in the same wave continue running.

Approval gates

If a connection has flow type Approval Gate, the workflow pauses when it reaches the gate. An approval card appears on the waiting agent, showing:

The upstream agent's output preview.
The approval prompt you wrote on the connection.
The downstream agent that's waiting.

Click Approve to proceed or Reject to skip the downstream agent (the rest of the workflow continues). If you don't respond within the timeout (5 minutes by default), the gate counts as rejected rather than stalling the run forever.

Ask-user prompts

If an agent calls the ask_user tool, a question panel appears on that agent's card. It shows the question and either a list of choice buttons or a free-form text input.

The agent pauses until you respond. Other agents (in the same or other waves) continue running.

Run cost

The status bar shows running cost. The Execution Modal breaks it down per provider, per agent. Costs come from the provider's reported usage:

Anthropic: reported tokens × per-model pricing.
OpenAI-compatible: reported tokens × pricing if your provider returns it (Ollama returns no cost; local = always $0).
Codex CLI: always $0 (subscription billing).
Cursor Agent: always $0 (subscription billing).

After a run, the full cost breakdown is saved to the audit log.

Audit log

Every run writes an NDJSON file to ~/.cerevisor/audit-logs/<session-id>.ndjson. The log contains:

Every event the orchestrator emitted.
Every tool call (name, arguments, result).
Every permission decision.
Every model call (provider, model, tokens, cost).
Errors with stack traces.

You can open these files in any text editor. They're the truth of what happened.

Replaying and partial re-runs

A .cerevisor file is the workflow definition. Re-opening it and pressing Run re-runs from scratch.

For partial re-runs after a completed run:

Resume from this agent — on a finished agent (in its log window, or the button on the card), re-runs that agent and everything downstream of it while keeping every other agent's output from the last run. A wave header has the same action for a whole wave. By default a confirmation preview opens first; a per-workflow setting can skip it.
Steer and re-run one agent — open a finished agent's log window and type in its chat box. Your message is treated as the latest instruction and just that agent re-runs with it (the canvas configuration itself is not changed).

Crash recovery

If Cerevisor crashes or is force-closed mid-run:

On next launch, Cerevisor finds the interrupted run and opens the Recovered interrupted work dialog.
All saved progress is restored and visible again — completed agents keep their output, turn counts, and cost. Interrupted agents are left stopped; nothing spends or acts until you say so.
For each interrupted agent you can click Resume agent to re-run just that agent (open its workflow first if it isn't active).
View recovered work opens the run view so you can inspect what finished. Discard deletes the recovered session.

From 3.2 Cerevisor also closes those runs out on its next start, so a run cut off days ago stops claiming to be running, and offers Continue this run — one click that keeps the agents that already finished and re-runs only what is left. See Continue an interrupted run.

Background runs

You don't have to keep the Execution Modal open. Close it (Esc) and the run continues. The status bar shows a live indicator with the current wave. Reopen the modal any time to see live state.

Cloud runs (Cursor only)

If you set the workflow's run mode to cloud and use the Cursor Agent provider, the run executes on Cursor's cloud VMs. Cloud runs survive Cerevisor restarts. Manage them via the Cloud Sessions modal (title bar icon). See Cursor cloud runs.

On your phone

The Companion shows the same picture of a run: its Overview (the counts, what it has spent, one tile per agent and any Cerevisor noticed lines), the Why believe this evidence behind a note when you tap it, the run's check by hand rows with the same two buttons, and any Question for you an agent has addressed to you — you can answer it from there. Reading needs nothing extra; marking a step by hand needs the same remote-control permission every other action from the phone needs, and a run that is not in the workspace your phone can see is not shown at all.

An older Companion app keeps working exactly as it did — it simply does not offer the new screens. Nothing on your phone needs updating for the rest of 3.2 to work.

See Pair and use the mobile companion.