Skip to main content

Run and monitor agents

Agents let Lotti carry a defined responsibility across multiple wake cycles. Each running instance comes from an agent template, uses an inference setup, and keeps its own state, reports, observations, and activity history. Open Settings → Agents to see the system as a whole.

Agent reports, observations, and queued wakes describe the agent's working state; they do not silently rewrite the personal logbook. When a task agent wants to change task data, Lotti presents a proposal for your review. See Review proposed changes for the per-task approval controls and their narrow initial-title and language exceptions.

Agent token usage is reported together with all other AI work in Review AI usage and impact.

Inspect the running instances​

Open Instances to see every agent and evolution session. Rows expose the agent kind, lifecycle, assigned template, and soul. Grouping by soul is useful when several responsibilities share one voice and decision style.

All screenshots
The Habitat Seal Inspector, Project Waddle Planner, and Sardine Cargo Coordinator remain separate instances even when they share system-level resources.
ScreenshotMobile view
Lotti agent instances grouped by Admiral Pebble, Dr. Flipper, Captain Sardina, and an evolution session

Select an instance to inspect its internals. The Stats tab shows model usage, its template, current inference route, lifecycle controls, wake count, and scheduling state.

All screenshots
The instance keeps enough provenance to explain which template and route produced its work before you pause, re-analyze, or destroy it.
ScreenshotMobile view
Habitat Seal Inspector instance detail with token usage, Orbital Habitat Sentinel template, inference route, lifecycle controls, and wake state

The remaining instance tabs preserve different evidence:

  • Reports contains the agent's saved report snapshots.
  • Conversations groups messages by wake cycle.
  • Observations exposes captured findings without mixing them into reports.
  • Activity shows the chronological message and tool trail.

Pause keeps the instance but stops normal work. Re-analyze asks a task agent to reconsider its current subject. Destroy is destructive and should only be used when the instance itself is no longer wanted; changing the template for future agents does not require destroying unrelated instances.

Understand wake cycles​

Open Wake Cycles to see work that is running now and work due later. A Pending wake was requested by the agent. A Scheduled wake was placed by the system or planner. Each row keeps the linked task, project, or day visible alongside the responsible agent and a live countdown.

All screenshots
The queue distinguishes an agent-requested habitat follow-up from launch-day and cargo schedules, so a countdown always has a reason and owner.
ScreenshotMobile view
Lotti Wake Cycles showing the running habitat inspection and scheduled Project Waddle and sardine logistics wakes

Deleting a queued wake removes that timer; it does not delete the agent or its linked task. If a wake immediately returns, inspect the agent state or the automation that requested it instead of repeatedly deleting the symptom.

Continue with Agent templates and souls to define what current instances do and how they communicate.