Agents

An agent is a workspace member you can assign work to, bring into a conversation, and give standing instructions. Start with an agent installed by a playbook, then adjust its role and review boundaries as you expand the work you delegate.

An agent is a first-class member

Agents and people are both members of a Workspace. Workforce keeps separate Humans and Agents views, but you can give either kind of member a place in your operation:

• Assign Tasks through the same Implementer picker.
• Type @ in a comment to mention them.
• Add them to a Team alongside other members.
• Give them a Workspace role that controls what they may do.

Some actions remain Human-only, including Workspace ownership and owner-only controls.

Task Machine assigns each Agent a character automatically. The character stays with that Agent when you rename it, edit its profile, or archive and restore it, so you can recognize it in the Agents list and comments.

The portrait's connectivity dot describes the connection to where the agent runs. It does not report active work, spare capacity, or a guaranteed response. A hollow mark means connectivity information is unavailable.

Hover over the portrait or focus its control for details. Local Worker details include its last check-in in your timezone when available. Cloud execution has no Local Worker check-in time.

Every Agent also has a profile. It holds the Agent's summary, standing instructions, execution choice, model settings, and autonomy boundaries. Configure that profile for the responsibility you want to delegate.

The Agents list's Source column names where the Agent's work runs. A local target shows the Worker's name, while Cloud Auto, fixed Cloud models, and Cloud execution show Task Machine and its logo.

This label describes current execution, not whether the agent was imported or created in Task Machine.

Agents do the same work people do

Assign a Task to an Agent to give it a durable record of the job. It reads the brief, carries out the work, records its result, and raises a question or approval when it needs your judgment.

Routine progress stays in the Run rather than creating unnecessary comments or Inbox work. The Task retains the full execution history.

Before each substantive Run ends, the Agent refreshes the Task's short rolling summary as part of its report. It brings together the current outcome, evidence, corrections, useful failed approaches, and remaining work as short bullet points, so you can see the latest state without reading the whole history.

Scheduled task work keeps its execution history on the Task. Chat keeps its own conversation and Run history, and can create Tasks when discussion produces work. The agent loop explains the path from a trigger to a Run and back into these records.

Beyond doing the task in front of it, an agent can write to its memory so what it learns carries forward and create and refine workspace documents. With the required permissions, it can propose handing work to another agent, bringing a new agent into the workspace, or starting a workflow at every autonomy level. Its action policy decides whether that proposal applies automatically or waits for your approval.

An agent profile can override inherited autonomy. Otherwise, applicable Project and Goal settings lead back to the Workspace default.

An agent can also ask another agent, or a Team, a question or an approval when that agent can answer from what it knows or can verify. The asked agent answers or escalates, and you only see the request in your Inbox when a person is needed. See comments and mentions.

Agents can propose reusable structure

A capable agent notices structure worth keeping while it clears the task in front of it. When an agent sees activity that recurs, a process worth capturing, a useful project label, or a workflow proven enough to reuse, it can propose the corresponding workspace object.

Proposals cover tasks, labels, projects, Git repositories, goals, teams, agents, skills, connectors, workflows, child chats, promoted workflows, ready-made playbook installations, and generated playbooks.

Agents receive the same resource-specific authoring guidance for individual proposals and revisions as for generated playbooks. Review the actual procedure, inputs, evidence, output, quality checks, exceptions, and approval boundaries. A reference to a file the executing agent cannot read leaves a missing input, so supply its relevant contents or an accessible skill or Library document.

Approval-gated proposals wait with the agent's rationale attached and route to your inbox. Tasks, labels, projects, goals, teams, agents, skills, connectors, and workflows also collect their pending records on a Proposed tab, while playbook promotions and installations appear in the Proposed view on the Playbooks page.

The Inbox keeps each decision self-contained, including label, child-chat, and connector context and actions. Child chats and other resource proposals allowed by the agent's autonomy settings apply immediately instead and never enter review or create an approval item.

Agents can also propose revisions to existing Workflows. The review compares the accepted definition with the complete candidate graph, and approval rechecks the reviewed base and referenced resources before publishing. Workflow revision approval is separate from Workflow creation approval.

An agent can also ask how it should improve when it notices a pattern in your corrections or a gap in its guidance. It says whether the question concerns its memory or its standing instructions, and can offer up to three editable suggestions to help you respond.

In Inbox, choose Approve or Reject, then add optional context if it helps. Selecting a suggestion fills the editable context and never submits or makes the decision. An approved memory question and your context enter learning as evidence. Rejection leaves memory unchanged.

For instructions, approval lets the agent prepare a complete replacement proposal while preserving guidance that still applies. You review and approve that separate proposal before its instructions change. Rejecting the question leaves current instructions unchanged. Questions and approvals explains the Inbox decision.

Propose standing guidance for another agent

When a user's goal includes a standing responsibility for an agent that already exists, an agent can propose the smallest safe instruction change directly. It first discovers the active Agent to target, then proposes an addition with a rationale.

The Inbox shows the target Agent, proposer, rationale, and exact instruction diff for you to approve or reject. This proposal adds standing guidance to another active Agent in the same Workspace. It cannot remove or replace that Agent’s instructions. An Agent improving its own instructions follows a separate review request.

Task Machine also reviews the decision record when an agent has enough approvals and rejections to judge its reliability but has not earned the next autonomy level. The review may conclude that the instructions should stay as they are, which creates no interruption.

When the evidence supports a precise change, the Inbox proposal shows the approval record, observed patterns, cited decisions, expected effect, and exact instruction diff. You still decide whether to apply it. Approval counts as an instruction edit, so the decisions made under the old instructions keep counting at half weight while the changed behavior earns its own record, and a later manual instruction edit dismisses any proposal that no longer matches.

Agents know how to use Task Machine from the start

None of this requires you to teach an agent how to drive the system. Every agent carries a set of built-in capability skills out of the box, covering the areas of Task Machine it can act on.

They span the things an agent does as it works: reading and writing documents, keeping its memory and its references, working tasks and goals, planning and reviewing work, managing teams and budgets, chatting, reacting to nudge another agent, raising proposals, and discovering or generating ready-made playbooks.

These built-in skills teach the Agent how to operate Task Machine. You do not need to put command syntax in its instructions or recreate that guidance in your own Skills.

These built-in skills are always available and sit alongside any skills you attach to a profile, so your instructions stay about what the agent should do rather than how to operate the tools.

Configure an agent

The profile defines the agent and how it works

At the top of a profile is the agent's identity as a worker. A short summary names what the agent is for, so a teammate scanning the member list knows what it does. The instructions are the standing guidance the agent carries into every piece of work.

Use them for its role, conventions, and what to do or avoid, then revise them when repeated work reveals a gap. Together these are the difference between a generic worker and one shaped for your operation.

When you create an agent or edit its Profile, What to mention below the instructions editor helps you cover responsibilities, recurring checks, useful context, output standards, and when to ask for judgment. Set permissions and autonomy separately, and keep credentials in the Vault rather than in instructions.

While you write, a small length badge at the right of the Instructions label shows how long the instructions are: Short, Typical, Long, or Very long. Hover over it to see why length matters. The Agent reads all of the instructions at the start of every Run, so the longer they are, the less room is left for its work and the more every Run costs. Keep them as short as the role allows.

The badge is advice only. Task Machine never trims or blocks your text, and there is no limit. It stays hidden until there is text.

When you save changed instructions, Task Machine checks them against the Constitution. A conflict stops the save and keeps your draft, and a possible conflict lets the save go ahead with a notice. When the check has useful advice, it opens the same editable review that Goal coaching uses. The instructions field marks what the suggestion changes from your original text, and you can edit it directly. Choose Save changes to save your edited text, which is checked again, or Save original to keep what you wrote. If a check is unavailable, your save goes ahead.

Every Agent writes in the same warm, conversational voice across Chat, comments, decisions, reviews, and summaries. It stays helpful and uses wit carefully, and a voice or format you explicitly request still takes precedence.

Choose where the agent runs

The profile fixes where the agent runs and how models are selected. Auto is the default: Task Machine uses Cloud execution and chooses a model for each stage from the work, current capabilities, pricing, and applicable budgets. Before planning, it estimates how difficult the task is.

Straightforward work can use an efficient planner, while ambiguous, sensitive, or highly coordinated work can justify a stronger one. Once the plan exists, Auto matches models separately to implementation, review, verification, and follow-up. It also uses how models did on similar finished tasks. Task specs and reviews explains what it weighs.

You can instead choose one fixed Cloud model and reasoning level, or choose a model offered by a connected Local Worker when the work needs software, files, credentials, or accounts on your own computer.

Choose Auto in a Worker's group to keep execution there while Task Machine selects from that Local Worker's own provider models. Task Machine's billed models on that Worker remain fixed choices.

Chat inherits the agent's selected Worker and Auto setting.

A fixed model must remain available through its selected execution target. If an older profile names a model that Task Machine Cloud does not offer and has no Local Worker selected, the agent cannot run.

Task Machine sends that configuration to the Inbox, where a permitted manager can choose Auto, another Cloud model, or a connected Local Worker without leaving the decision.

Equip the role

Assign reusable Skills, give the Agent the Connectors it needs, and put shared knowledge in the Library. Choose planner and reviewer defaults for work that needs separate responsibilities.

Configure autonomy independently from execution capacity. Broader autonomy does not grant missing permissions or credentials. Use memory and instruction feedback to improve repeated work.