How Task Machine works

Task Machine coordinates work you delegate to people and agents. Start with an outcome, let the assigned team work, and handle the questions, decisions, and results that need your judgment. The same loop supports one job and work that repeats.

Operating loop

Choose the work. Describe a useful result, discuss it in Chat, or select a Playbook. Give the agent the inputs, constraints, and definition of success it needs.

Let the team run it. Tasks carry individual jobs. Workflows carry explicit repeatable steps. An agent may need clarification, a reviewed plan, access to a service, or your approval before it can continue.

Handle what needs you. Questions, approvals, proposals, exceptions, and reviewable results arrive in Inbox. Read the decision context and resolve the item there.

Steer the details. Use the Task when a particular job needs direction or its history needs inspection. You can correct the work without turning Chat into a status board or monitoring every Run.

Repeat proven work. Once you trust the process, add schedules, appropriate autonomy, budgets, and verification. Learned autonomy uses your recorded decisions to recommend changes in delegation, with each change left for you to approve. These controls support delegation without guaranteeing correct output.

Use three surfaces for daily work

Chat to decide, Inbox to approve, Tasks to steer is the short version of the product.

Chat is for open-ended discussion, planning, and creating work from that discussion. It keeps its own conversation and execution history.

Inbox collects the moments that need you. A useful decision includes its context, consequences, and resolution controls. Inbox is where you should spend most of your day-to-day attention after setup.

Tasks hold individual jobs and their durable history. Activity carries discussion and steering, while execution details remain available in Runs. Open a Task deliberately when you need that detail, rather than keeping every Task open to watch its agent.

Follow one report from discussion to repeatable work

A weekly client status report shows how the surfaces and records work together. The report needs account activity, delivery progress, open risks, and your judgment before it goes to the client.

1. Describe the outcome in Chat. Explain who the report is for, which sources the agent should use, what claims need evidence, and what must remain for your approval. Ask Task Machine to turn that discussion into a repeatable process.
2. Review the proposed plan in Chat. The plan can create a Task for the first report and a Workflow that collects updates, drafts the report, verifies the evidence, and pauses for approval. Correct the steps or responsibilities before you approve the proposal.
3. Let the first Task run. The assigned agent follows the approved Work Spec and Workflow. If account activity is missing or two sources disagree, the agent records the gap instead of inventing an update.
4. Answer the exception in Inbox. The question arrives with the affected report, the missing context, and the actions needed to continue. Answer it there, and the same Task resumes without a separate status conversation.
5. Review the result in Inbox. A verifier checks the draft and its evidence before the approval reaches you. The approval item carries the report, unresolved risks, and the choice to approve or return it for revision. Approval accepts that reviewed result. It does not send or publish anything unless the Workflow includes and authorizes that action.
6. Use the Task when you need detail. Its activity and Runs preserve the plan, questions, corrections, verifier result, and final outcome. Once the handoff works reliably, schedule the Workflow for the next report rather than rebuilding the instructions each week.

This is the operating loop in practice: decide in Chat, handle judgment in Inbox, and open the Task when a specific job needs steering or inspection.

Distinguish the job, the worker, and the process

A Task describes one piece of work. An Agent is a workspace member that can be assigned work and use the context and capabilities it has been given. A Run is one execution attempt, not necessarily a completed Task. A Task can need several Runs as it receives answers, corrections, and reviews.

A Workflow is an explicit sequence of steps with configured branches and checkpoints. A Playbook installs a setup for a job, which can include agents, documents, goals, and workflows. Start with a Playbook when one fits, then learn the Workflow builder when you need to change the process.

The agent loop connects product events to execution. Assignments, schedules, Chat messages, and conversational steering can start work. Its results update product records and can lead to the next action. A schedule becoming due, an agent finishing a Run, and a human approving a plan are different events with different consequences.

None means every remaining action is already authorized.

Add structure when the work needs it

A Workspace holds the people, agents, work, settings, and access boundaries for one operating context. Within it, Projects group related work and Goals describe outcomes. Teams group members. A solo user does not need to create an organization chart before delegating a first job.

The Library holds reusable documents, while Memories carry notes an agent can use in later work. Connectors provide access to external services, and Vault stores credentials. Add the specific context and access a job requires rather than configuring every available service.

Each supporting concept has its own guide. The quickstart introduces only what the first job needs.

Choose where agents run

Cloud execution is the default and works without a connected computer. A Local Worker lets agents use supported tools and resources on a computer you control. A Runtime is an installed agent tool available on that Worker.

Cloud work consumes workspace usage. Local work uses your own tools and provider arrangements. Workspace pricing and access explains the financial boundaries, and Local workers explains execution choices.

Keep judgment and access explicit

A workspace role controls access and actions. An agent's effective autonomy controls which supported actions it can take without another approval, including inherited settings and any applicable overrides. Budgets, required reviews, verification, and service access place additional boundaries around execution.

An agent can propose follow-up work without that proposal being approved. A verifier can report a pass without supplying a required human sign-off. A completed Run can still leave the Task waiting for an answer or review. Read the current state and the requested decision rather than treating all progress signals as completion.

Questions and approvals explains how to handle those decisions. Start with Get your first result, then use Inbox to operate the work that follows.