Spools

Spools: follow runs and review approvals

Send tasks to an agent, understand queued and paused runs, review MCP calls, stop work, and pass a canvas node without sending the whole workspace.

Before you start

Open Spools through the account menu → Agents and select an agent. Spools is in early access and only appears for accounts that have been given access; upgrading your plan does not enable it.

Check Agent settings before sending a task. Tool availability depends on its switches, the model, and the relevant services. Connected apps (MCP) uses your enabled account connections from Settings → Connected apps; the current canvas workspace’s MCP switches are not the agent’s connection selector.

An approval is permission for a particular tool call, not confirmation that its arguments are correct or that the whole task is safe. Integrated computer and canvas tools have different rules from external MCP tools.

When to use this

Use the agent chat to give a specific task, follow its progress, and review any proposed external changes. Use Send to an agent when the task should start from an existing canvas answer or note rather than a manually pasted transcript.

For changes to external records, identify the exact resource and desired outcome. Ask for a draft first when the action could send, publish, spend money, or alter important information. Written constraints help communicate intent but are not a technical approval gate for every integrated tool.

Steps

Select the agent and write a clear task in its message box. Use the send button or Enter; Shift+Enter inserts a new line. Specify the scope, expected output, and actions it must not take.
Watch Waiting for its turn… while the run is queued, then Working… during execution. These messages do not guarantee a start or completion time. Work is processed in the background; closing the tab is not the same as stopping the task.
Review tool cards. Expand Arguments to see the actual proposed operation and Result after it runs. In Spools, MCP tools classified as read-only run automatically. Other MCP calls, including writes and destructive calls, pause the run at Waiting for your approval. Classification relies on the tool’s metadata; do not treat a read-only label as independent proof of safety.
Choose Allow or Deny for each pending call. If several are pending, the interface collects your choices and submits them once all have a decision. Allow all approves every currently shown pending call; inspect all of them before using it. There is no agent-specific Always allow choice in these cards.
After decisions are submitted, the run is queued again and continues with approved results or denied-call information. Denying one call does not automatically stop the whole task: the model may continue without it or propose another call.
To stop a queued or running task, use Stop in the composer; for a paused task, use Stop beside the approval message. A running task may need time to notice the request, and an already-started operation may still finish. Check the resulting status and any external changes.
To hand over canvas context, open a node’s menu → Send to an agent, or use that action on a supported answer or note in Chat. Select the agent, edit What should they do with it?, and select Send. The agent must have Neurothread canvas enabled; the dialog blocks sending when it is off.
After Failed or Interrupted, inspect the text, tool results, and external service before trying again. Correct settings if needed, then send a new, narrower message from the agent composer. There is no dedicated retry or restore-run button in this chat.

Worked example

Illustrative exercise, not an executed run: from a project-plan node, select Send to an agent and ask: “Read the conversation leading to this node and identify three missing acceptance criteria. Do not create tasks yet.” The dialog sends your task plus the node title, workspace name, and node ID as a reference; it does not dump every workspace into the message.

If you later ask the agent to create one tracker task through MCP, inspect the project, title, and description under Arguments. Allow only the intended call. If two calls are proposed and only one is appropriate, allow one and deny the other before the run resumes.

What happens next

The conversation retains run text, tool statuses, errors, and recorded charges. States include queued, running, awaiting approval, succeeded, failed, cancelled, and interrupted. Successful completion means the run finished, not that every factual claim or external change has been independently verified. Reopen the agent to inspect saved progress or the final result.

The model receives a bounded recent history of successful turns, not the entire lifetime of the agent conversation. Canvas tools can search your owned workspaces, read the ancestor conversation of the referenced node, or inspect a workspace tree when requested; reads can be truncated. A node handoff is a starting reference, not a sandbox that prevents other canvas access. Turn the canvas tool off when that broader capability is unnecessary.

Integrated memory, canvas writes, and computer operations can run without the MCP approval pause. Stop does not undo a sent message, created record, saved canvas node, file change, or remembered note. Interrupted runs are not automatically replayed because tools may already have had effects.

Cost and limitations

Model work, searches, and computer use can consume credits. A pause, denial, failure, or stop does not refund work already performed. Resuming after approval can incur additional model and tool costs; the saved run accumulates those charges. BYOK applies to model work only when the account has saved a key matching the selected provider model and its plan permits BYOK. Neurothread house models, including the default agent model, still use platform credits. There is no separate agent key setting. BYOK does not remove computer or applicable search charges; web search is disabled when BYOK is active and the account is out of credits.

Tool results and conversation history are bounded. The short Result saved in a card is not necessarily the complete external response. Missing live updates do not prove that a task has stopped. No fixed queue wait, run duration, or immediate cancellation is guaranteed.

Troubleshooting

Waiting for its turn for a long time: the task is queued. Do not send duplicates just to force it to start; reopen the conversation to check its state.
Approval selections do not resume the task: decide every pending call, not just the first one. A stale approval may report that the task is no longer waiting; reopen the conversation.
The agent reads an app without asking: read-only MCP tools are automatic in Spools. Disable Connected apps (MCP) in agent settings if you do not want that access.
A canvas or computer write ran without a card: integrated tools run automatically. The MCP approval rule is not a universal write-action barrier.
Send is disabled in the handoff dialog: create an agent first, select one with canvas enabled, and write a non-empty task.
Interrupted or failed: inspect partial effects before sending again. State what was already completed; a new message is a new run and may incur new charges.
Stopped but something changed: stopping cannot reverse a completed or already-started operation. Review and repair the change in the affected service if appropriate.