The deep-agent brain

Most questions deserve a fast, direct answer, and that is what Harness gives by default. But some questions are harder: they have several parts, they need real digging, or the answer has to be right rather than plausible. For those, Harness has a deeper reasoning mode.

What it does

On a hard question, the deep-agent brain does the extra work a careful person would:

• Plans the steps instead of answering in one shot.
• Gathers from the sources that matter (the web, your screen, your memory) with focused lookups rather than one broad guess.
• Pursues each part of a multi-part question instead of stopping at the first result, running independent threads in parallel where it helps.
• Verifies the load-bearing facts before it commits to them, and keeps what it verified visibly separate from what it is inferring.

The result is an answer you can trust on the questions where trust matters, with its reasoning grounded in what it actually found this turn rather than assumption.

How it works

Harness runs a two-speed engine. Every turn is classified as fast or deep:

• The fast path is the default and stays byte-for-byte what it always was: an instant, screen-grounded answer with light tool use. Everyday questions never pay for depth they do not need.
• The deep path engages only when a turn genuinely calls for it. It routes to a stronger frontier model, turns up reasoning effort where the model supports it, plans the work, and can fan independent sub-questions out to research subagents that run in parallel and report back. Load-bearing claims are run through a verification step, and anything that does not hold up is dropped or hedged rather than asserted.

This split is deliberate: the fast path guarantees the everyday experience never regresses, while the deep path adds depth as a separate, additive lane on top of it.

Availability

The deep-agent brain is a Resident capability, and it is server-authoritative: whether a turn takes the deep path is decided on the server against your entitlement, not claimed by the client. See Plans.