Agents and automation

One session can involve more than one agent: the one you talk to, the sub-agents it hands work to, and the steps of a workflow you approved. Each one works under the same permissions, and each one is visible.

Work mode: normally, or plan first

A session works in one of two modes, chosen in its permissions panel, next to the permission:

• Work normally — the agent acts as it goes, within its permissions.
• Plan first — the agent prepares a plan and takes no action. When the plan is ready it presents it and the turn pauses until you choose: proceed asking for confirmations, proceed accepting edits, proceed in a clean conversation, or keep planning with your feedback.

The agent can also ask you to switch to Plan first when a task needs a plan before any change; the switch happens between turns, and a banner in the chat says so. When it needs a decision from you, it asks a question with options.

Sub-agents

The agent can hand a self-contained piece of work to a child session: it starts with nothing but the instruction it was given, works on its own — reads, tries, fails, tries again — and reports back only a final summary. None of its intermediate steps enter the main conversation; that is the point of it. By default it works in the same folder.

You do not start a child by hand: the agent decides to delegate when the work divides well. You follow the children in the Agents tab of the right panel: for each one its state (running, done, failed, interrupted), its model, its turns and its tool calls. A child’s conversation opens inside that tab, with a way back to the list; the parent’s chat does not move.

• At most 10 children at once under the same parent. Beyond that a delegation is refused, and the refusal says why.
• At most two levels deep: a child can have children, and there the chain stops.
• A child’s model shows as a dash until the child declares it — a fallback name would be an invented model.

Workflows

For larger work the agent can propose a workflow: a draft in phases and steps that the server checks before you see it. Proposing approves nothing and starts nothing: you review it, approve it, and start it; then you follow its run step by step, pause, resume or cancel it, and, once it has finished, retry its failed steps. The operations behind it are in Workflows.

The Board

The Board is the table of every session of TALOS Desktop: session, state, model, turns, tokens, cache, time to the first token, how it ended, and when it started. Usage columns count the whole session, not its last message. A usage that was not measured is a dash, never a zero, and an old session with no recorded outcome reads “done · outcome not recorded” — which is not “succeeded”. Sort it by column, and filter it.

Automations

An automation starts a task again by itself, every so many minutes, with a cap on runs per day. Automations in the sidebar shows, for each one, its interval, whether it is active or paused, how many times it ran today out of the cap, and when it last ran and runs next.

It records the schedule and the starts, not outcomes or costs: what happened in a run is read in its session. A paused automation does not catch up on the runs it missed, and when the daily cap is reached the next run reads “daily limit reached” instead of a false date. See Schedule an automation.