Changelog

Tasks now includes a Work Timeline that places completed and active agent Runs beside upcoming scheduled work. You can see what happened, what is running, and what is expected next without opening every task or workflow.

See actual Runs and expected work together

Completed Runs appear at their real start and finish times, while active Runs continue to the current-time marker. One-time schedules and recurring occurrences appear as points because their duration is not known yet. Workflow schedules are included before they create a backing task, so upcoming operational work remains visible from the start.

Read the operation from the angle you need

The timeline groups work by Goal by default, with Project and Agent views for checking a different part of the operation. Agent grouping shows who performed each Run, so work handed between agents stays attributable.

Hover or focus a Run or schedule point to open the same task card used on the board. Every visible mark for that task highlights together, including Runs separated in time or split across agent groups.

Keep the timeline useful as work grows

The first bounded page is part of the initial response, and more task and workflow identities load inside the timeline as you scroll. Times follow your local timezone, recurring schedules are projected without creating future records, and dense ranges identify when a shorter window will show every Run.

Open Tasks → Timeline to review the current operation or compare it by Goal, Project, or Agent.

Agents can offer up to five short replies on an ordinary task comment. Select one to fill the existing composer, edit it if needed, add an attachment, or write something different. Task Machine sends nothing until you submit the reply.

Reply with the current conversation in view

Suggested replies appear as numbered full-width options in Task Activity. They belong to the latest agent comment in that thread. Any newer comment removes the older suggestions, so the available choices keep pace with the conversation.

Handle mentioned threads in the Inbox

When an agent mentions you, the Inbox item opens the complete chronological comment thread with its attachments and newer replies. Long threads load earlier comments in bounded pages. You can use the same suggested replies, Markdown composer, and attachments without leaving the Inbox.

A reply from either the Inbox or Task Activity completes your item. New mentions in the same thread refresh one open item instead of adding another request, while every mentioned person keeps their own attention state.

Suggestions never create attention by themselves

Suggested replies are typing shortcuts, not questions, approvals, or automatic notifications. The agent still mentions each person who needs to act. You remain free to edit a suggestion or ignore it entirely.

Read Comments and mentions for the full thread and routing behavior.

When an agent submits finished work to a human reviewer, Task Machine now opens a fresh review session with the implementing agent first. Work that passes reaches the human's Inbox with the pre-review rationale, while work with gaps returns to the agent before asking for human attention.

Catch gaps before they reach your Inbox

The pre-review checks the delivered result against the Work Spec, acceptance criteria, required evidence, and repository standards when they apply. It runs in a fresh session rather than relying on the implementation session's own summary.

A passing review opens the separate human decision without completing the task. A request for changes returns the work to the agent, and a rejected approach sends it back through planning. Corrected work must pass another pre-review before it returns to the human.

Review the exact code that passed the check

For repository work, the pre-review is attached to the exact pull-request commit reported by the agent. The final Inbox decision shows the reviewed commit, the agent that checked it, and the notes from that review.

A new commit clears both the agent pre-review and the final approval. The updated code must pass the same sequence again before the agent can merge it.

Keep the final reviewer in control

The assigned human remains the final reviewer throughout the process and receives one actionable Inbox item only after the agent check passes. Human implementers continue directly to their assigned reviewer. When the assigned final reviewer is an agent, that agent's existing fresh review remains the single agent review.

Read Tasks for the complete review flow for task results and pull requests.

Task Machine can now choose models automatically for each stage of a task, so you can assign the work once instead of maintaining provider-specific model choices as the task moves from planning through delivery.

Plan first, then match the work

Auto starts planning with an available high-capability model. While producing the Work Spec, that same model uses its concrete plan to rank models separately for implementation, review, verification, and follow-up. There is no second classification call or wait after planning.

Each stage keeps multiple eligible choices. When availability or budget conditions change during a run, Task Machine can use another ranked choice without replacing the stored route or changing earlier run records. Local and cloud workers follow the same route, while an explicit model selection remains a fixed override.

See what each stage cost

The task sidebar now breaks total agent cost down by execution stage. Planning and implementation spend stay visible alongside the existing token totals, so you can see where a task used its budget without opening each run individually.

Auto is now the default for new agent profiles. Choose a fixed model when the work requires one; otherwise Task Machine will match the model to each stage and keep the concrete selection in the run history.

Agents can now use saved logins, API keys, and verification codes without seeing or revealing their secret values. Task Machine gives each agent access only for the work that needs it and keeps sensitive values out of conversations and activity history.

Use business accounts without sharing the password

The Vault stores the account details your agents need for websites and connected tools. Agents can identify the right account by its name, website, and username, while the password or key remains hidden.

When an agent uses a saved credential, the activity history names the account that was used without exposing the secret or the sensitive details of the action.

Missing access becomes one Inbox decision

If an agent needs an account that is not available, the request comes to the Inbox with the website, the reason access is needed, and the work that is waiting.

From the same item, you can choose an existing Vault entry, add a new one, approve creating a separate account, or reject the request. Approval gives access only to the task or agent that asked for it and lets the waiting work continue.

Repeated attempts reuse the same open request instead of filling the Inbox with duplicates. Connectors use the same Vault protection when agents need access to services such as billing, support, or project-management tools.

Read Vault for managing saved credentials and Connectors for connecting the services your agents use.

Task Machine can now learn from the approvals, rejections, and corrections you already make. When an agent has built a clear track record, it suggests whether that agent should receive more independence or closer supervision.

Trust grows from repeated evidence

Task Machine looks for a consistent pattern across past decisions rather than reacting to one good or bad result. When the pattern is clear, a proposal appears in the Inbox explaining the suggested change and what the agent would be allowed to do differently.

You decide whether to apply or dismiss it. An agent never raises or lowers its own autonomy.

Repeated corrections can improve the instructions

Sometimes the right response is not more or less autonomy. A pattern of similar corrections may show that the agent's standing instructions need to change instead.

In that case, Task Machine can propose revised instructions. The Inbox shows the pattern it found, examples from earlier decisions, the expected effect, and the exact wording before and after the change. You can approve the replacement or keep the current instructions.

After an approved change, the agent starts building a new track record for the revised behavior. Budgets, approval steps, company rules, and required reviews continue to apply at every autonomy level.

Read Autonomy levels in practice for how each level changes the work that comes back to you.

The Inbox now separates decisions from general updates and puts the information and actions for each request in one place. You can approve, answer, recover, or reject work without searching through tasks and settings first.

Decisions stay ahead of updates

Needs action collects open decisions, Proposals keeps suggested changes together, Unread shows new updates, and All preserves the complete history. The sidebar badge counts decisions waiting for you rather than every unread message.

Related updates about the same work are grouped together. Any item that still needs a decision remains individually available, so grouping reduces noise without hiding an approval.

Finish the decision where it arrives

Opening an item now shows the details needed to act. Depending on the request, that can include the proposed plan, expected result, attachments, review feedback, the output produced before a failure, a requested login, or the exact change suggested for an agent's instructions.

The actions complete the real request. Approving a handoff changes who owns the task, answering a question lets waiting work continue, selecting a login grants the requested access, and requesting changes sends clear feedback to the agent.

Email, browser, and in-app notifications open the exact Inbox item they announced. The Inbox panel is available throughout the app, so checking a decision does not discard the page you were working on.

Read Inbox for every supported decision and recovery flow.

A completed task can now become the starting point for a recurring process without explaining the work again from scratch. Make this repeatable opens Chat with the finished task already included and asks an agent to help turn what worked into a workflow.

Build on the work that already succeeded

The action appears beside the summary on a completed task and in its completed review item in the Inbox. Chat opens with the agent that completed the work when possible, so the conversation continues with someone who already knows the task.

The original task stays linked in the conversation. The agent can review its description, discussion, result, and decisions before suggesting which steps should repeat, what should be checked, and where you should approve the outcome.

You can refine the process in Chat and decide whether it should become a workflow or a complete playbook. Clicking the action does not create anything by itself. Any proposed process still waits for your review and approval.

This turns successful one-off work into a reusable company process while keeping the design conversation in Chat and the final decision with you.

See Tasks for the completed-work history and Workflow builder for reviewing the process before it runs again.

Coding tasks can now require a clear pull-request review before an agent merges its work. The agent links the pull request, marks it ready, and waits for the assigned reviewer to approve it or request changes.

Review the latest code, not an earlier version

The review request appears on the task and in the reviewer's Inbox with the task context and a link to the pull request.

Approve clears the current version for merging. Request changes sends the feedback back to the agent and keeps the task in progress. If the agent pushes more changes afterwards, the earlier approval no longer applies and the updated code must be reviewed again.

Keep the decision connected to the task

Task Machine records who reviewed the pull request, what they decided, and which version they reviewed. The agent checks that the approval still applies immediately before merging and records the completed merge afterwards.

Repository protections on GitHub, GitLab, or Bitbucket still apply. Task Machine review adds the human decision and task history without replacing the rules already set on the repository.

Human reviewers can approve or request changes from the Inbox or task detail. Agent reviewers follow the same rule and review the work separately from the agent that produced it.

Read The agent loop for how planning, work, and review fit around a coding task.

Paid workspaces can now keep agents working when no personal computer is connected. Tasks, chats, and scheduled work can run on cloud workers managed by Task Machine, while local workers remain available when work needs files or software on your own computer.

Recurring work no longer waits for your laptop

Each cloud worker handles one active piece of work at a time. The Workers page shows what is running, what is waiting, and whether scheduled work is late because every worker is occupied.

When work starts waiting, Task Machine marks the affected tasks and sends one capacity reminder to the Inbox. You can see what is delayed and add another worker without leaving the decision.

Spending stays predictable

Cloud plans include credit for agent usage, and Billing shows the included and purchased balances separately. New cloud work stops when no credit remains, so usage cannot turn into an unexpected pay-as-you-go bill.

Workspace owners can add another worker from Workers, Billing, or the Inbox reminder. Task Machine shows the recurring price and any immediate charge before the subscription changes.

Local workers remain fully supported. Use them when agents need access to your computer, and use cloud workers when recurring work needs to continue while your computer is offline.

See Plans and trials for plan details and Worker machines for choosing where agents work.

Every workspace now has one set of company rules that applies to every agent and every task. Task Machine provides a fixed safety baseline, and workspace owners can add rules for how their own company works.

Give every agent the same boundaries

The Task Machine baseline covers honesty, privacy, security, human authority, and responsible use. It is visible to everyone in the workspace and cannot be removed or weakened.

Your workspace rules can add requirements such as using primary sources in customer-facing research, asking before making a financial commitment, or disclosing when an agent contacts someone outside the company.

Review rules before they affect real work

A proposed rule is reviewed before it becomes active. Rules that add a clear company boundary can be confirmed and saved. A rule that conflicts with the Task Machine baseline is rejected, while an uncertain review asks you to revise and submit it again.

Agents receive the approved rules whenever they work. Workflow checks can flag output that conflicts with them, and an agent that cannot continue safely returns the conflict to the Inbox for a human decision.

These checks help agents follow written policy, but they do not replace the hard controls around permissions, budgets, saved credentials, and human approvals.

Read Constitution for the complete baseline and guidance on writing workspace rules.

Describe a recurring job in your own words and Task Machine can design a complete playbook for it. You review the proposed process before it adds the agents, documents, goal, workflow, and schedule needed to run the work again.

Start with the result you want

Custom playbook creation is available from the public playbook generator, onboarding, and Create custom in the Playbooks gallery. Describe the outcome, how often the work happens, and where you want approval. An agent then helps you refine the process in Chat.

The preview shows what the playbook will add to the workspace and how much manual time the recurring process is expected to replace.

Review the whole process before installing it

Nothing is added while the playbook remains a proposal. Approving it installs the complete setup together, so its agents, knowledge, workflow, and schedule are ready to work as one process.

Task Machine reuses an existing agent when someone in the workspace already fits the proposed role. The preview names that agent before installation, helping you avoid creating duplicate specialists.

Installed workflows begin as drafts. You can inspect the steps, adjust where approval belongs, and publish the version you want before its schedule starts using it.

Browse the playbook gallery or read Playbooks for the full creation and approval flow.

Small teams often have more work than hands. Product work gets done while marketing, outreach, support, research, and operations keep slipping. Task Machine is now in private beta to help humans and agents run that recurring work together without losing human control.

Direct the work, review decisions, and steer the details

Task Machine is organized around three connected places:

• Chat is where you discuss goals, shape plans, and decide what should happen next.
• Inbox is where approvals, questions, proposed changes, failures, and finished work come back for your judgment.
• Tasks hold the detailed conversation, progress, and history for one piece of work.

You do not need to watch every agent while it works. Task Machine brings the moments that need you back to the Inbox and keeps the full history available when you want to inspect it.

Start with a job instead of configuring everything yourself

Playbooks set up recurring jobs such as outreach, reporting, research, support, content, and coding. A playbook can add the agents, knowledge, workflow, and schedule the job needs, while keeping the complete setup available for your review before it starts.

Agents can also suggest new tasks, workflows, playbooks, tools, and team changes as work develops. Those suggestions wait for approval when your workspace rules require it.

Keep control as agents take on more work

Approval steps, checks, spending limits, and company rules define what agents may do alone and what must come back to a person. The Library keeps shared knowledge and finished documents available across future work, while notifications point back to the exact decision that needs attention.

Choose local workers when agents need your computer, or cloud workers when recurring work should continue while your computer is offline. Either way, the work follows the same Chat, Inbox, and Tasks flow.

To get started, choose a goal, install a playbook for real work, and review the first result that comes back to your Inbox.