Working depth, schedules and watches

Depth

You choose how hard Metis thinks about a question, per conversation, from the selector in the chat toolbar. There are four settings:

• Fast — quickest and cheapest, for everyday lookups.
• Standard — balanced. This is the default and the right answer most of the time.
• Deep — closer analysis, for document-heavy work.
• Max — the strongest reasoning available, for genuinely hard problems.

Depth changes which model runs, never what Metis is allowed to do. Every mode has the same tools, the same access to your data and the same guardrails. A deeper mode uses your allowance faster, which is the only reason not to leave it on Max permanently.

If the provider behind a mode is unavailable, the mode falls back automatically rather than failing.

On the free plan there is no selector — conversations run in Fast. See pricing.

Having it work while you are not there

Ask Metis to do something later, on a repeating schedule, or when something new appears:

"Every Monday morning, summarise what changed for my fleet last week."

"When a new Airworthiness Directive is published for the A320 family, assess it and tell me."

The first is a schedule; the second is a standing watch — it does not poll on a timetable, it reacts when a matching document lands.

Creating one needs your approval, and the preview shows you exactly what it parsed from your sentence before you agree, so a misread filter is caught before it runs rather than after.

Results reach you through the notification bell, and through Telegram if you have paired it. Your scheduled work is listed under Scheduled work in the sidebar, where you can see what is running and stop it.

The rules unattended runs follow

• Restricted tools and approval refusal. Unattended runs use an explicit allowed tool set. Actions requiring interactive approval are refused. Research can still use permitted external services and user material, and reports are delivered through the configured notification channels.
• Silence when there is nothing to say. A daily check that finds nothing writes to its own journal and sends no notification. Watches that cry wolf get muted by their owners, so they are built not to.
• A broken watch stops. Repeated failures notify you once and then pause the watch, rather than failing quietly forever or filling your bell with the same error.

Each watch keeps its own record of what it has already reported, so you are not told about the same document twice.

Check the run outcome and the report's coverage. Silence does not establish that nothing changed. Keep required official monitoring and a fallback for failed or late runs. Scheduled runs do not automatically load your general profile and working rules, but their permitted tools can recall relevant user material; one-shot background tasks can also carry the originating conversation.