Assign a task agent
A task can have one task agent. Assigning one gives that task a persistent AI assistant with its own inference setup, report, wake state, and proposal history. The agent reads the task and its linked context, distils the current situation into a short report, and can surface proposed task or checklist changes for you to confirm or reject. For the system-wide view across every running agent, see Run and monitor agents.
The assigned agent appears below the task's checklists and linked tasks. Its collapsed card keeps the latest summary, automatic-update control, Wake agent action, and active model route visible without making the task page feel like an AI console.
On a phone, the card follows a single reading order: identity, update controls, AI setup, report, then proposed changes and history. On desktop, the report and proposals form the reading column while the compact controls remain nearby.
Choose Read more to expand the complete report in place. The expanded card keeps the task visible, adds the report's supporting assessment and recommended next step, and offers Open agent internals when you need reports, conversations, observations, or runtime details.
Choose the agent's AI setup
The card's Current setup row identifies the profile, model, and provider used for its next wake. Select it to choose a saved inference profile, override only the thinking model, return to the profile default, copy the category's current setup, or turn off AI for this one agent. A choice stays with that agent for future updates; copying a category setup is a snapshot, not a link that silently changes later.
If a report was produced through a different route, Lotti keeps a separate This report attribution instead of relabelling history with the current setup. Turning AI off leaves the existing report readable but prevents further wakes until you select a working setup again. Automatic updates cannot be enabled while the agent has no working AI setup.
Review proposed changes
A completed wake can update the report and add proposed changes. Automatic updates decide when the agent wakes; they do not give the agent permission to edit the task. Proposals remain pending until you act on them.
Use the cross to Dismiss a proposal or the check to Confirm and apply it. You can also swipe a proposal left to dismiss or right to confirm. When several proposals are pending, Confirm all applies the currently listed changes together. Review bulk proposals carefully, especially status, date, and checklist changes.
Open History to revisit recent outcomes. It distinguishes confirmed changes from dismissed ones. A dismissed entry can represent either your decision or a proposal the agent retracted after newer task context made it stale.
A manual Wake agent can produce the same kinds of report and proposals as an automatic wake. Turning automatic updates off stops future background wakes; it does not apply, discard, or otherwise resolve proposals that are already waiting for your decision.
Choose automatic or manual updates
New task agents start with Automatic updates off. Assignment alone does not spend tokens in the background. While updates are off, Lotti continues to notice relevant changes to the task and its linked entries, marks an older summary as out of date, and waits for you to choose Wake agent. A manual wake remains available whenever the agent has a working AI setup.
The out-of-date notice deliberately leaves the last useful report in place. If a relevant edit arrives after a wake begins, that newer edit keeps the report marked stale even when the earlier wake finishes. When you need a report that includes those latest changes, wake the agent again after the task has settled.
When Automatic updates are on, Lotti bundles matching task changes for about two minutes before scheduling a refresh. This coalescing avoids an agent wake for every small edit. Wake agent is still available for an immediate, user-requested refresh.
Turning automatic updates off clears its pending countdown and queued automatic refreshes, but does not cancel a wake you explicitly requested. Turning it on again applies to future changes; it does not replay changes that arrived while automation was off.