Deep work: Ultra, research, and goals

Ask your workspace "what did we decide about pricing?" and the agent answers in about thirty seconds: one search, a few citations, done.

Some jobs are bigger than that. "Research the Korean shipbuilding industry and build me a sourced reading folder." "Merge every trip note into one clean itinerary." Those take dozens of tool calls and sustained attention, and approving every step by hand would eat your whole afternoon.

Consilience has three deep modes built for exactly this. Ultra fans out parallel read-only explorers for hard questions. /research turns the web into sourced documents inside your workspace. /goal keeps the agent working on its own until a condition you wrote comes true. Steering rides alongside all three, so you can redirect a long run instead of restarting it.

Ultra

The top notch of the intelligence dial. You get the best reasoning plus read-only sub-agents that explore in parallel.

/research

A slash command that deep-researches the web and saves one document per source into a research/ folder.

/goal

A slash command that keeps the agent working, turn after turn, until the condition you wrote comes true.

steering

Sending messages into a run that's already going. Enter queues yours for the agent's next step; ⌘Enter cuts in and delivers it now.

background task

A slow command the agent parks off to the side so it can keep working, shown as a "Running in background" chip.

Ultra: parallel explorers for hard questions

In the composer, the intelligence badge shows your current model and its effort level. Tap it and the "Intelligence" dropdown opens, where you set how hard the model thinks: "Off", "Minimal", "Low", "Medium", "High", "Extra High".

Above all of those sits one top notch, "Ultra", described as "Top reasoning plus read-only sub-agents that fan out and explore in parallel." Reach for it when one straight-line pass won't do and a question needs many corners of the workspace checked at once.

• Your message is routed through an orchestrator that launches parallel read-only sub-agents.
• Each explorer shows up as its own "Exploring…" chip that settles to "Explored", so you watch the search fan out in real time.
• The sub-agents are read-only on purpose. They investigate, and only the main agent writes anything.
• One press of the stop button cancels the whole thing cleanly.

/research: turn the web into workspace material

Type / at the start of a message and the slash picker offers "Research", labeled "Deep-research the web into a workspace". The shortcut ⌘⇧R (Ctrl+Shift+R) does the same from anywhere, opening a chat with /research already filled in. Give it a topic and the deep-research orchestrator takes over.

/research korean shipbuilding industry

That last step is the whole point. The sources are markdown documents, so they get extracted into your knowledge graph like anything else you write. Later questions can cite them, and the entities they mention join your graph.

/goal: keep working until it's done

The other deep slash command is "Goal", labeled "Keep working until a condition is met". You describe an end state, and the agent loops on its own, turn after turn, checking the condition each round. It keeps going until the condition is met, it hits the turn cap, or it gets stuck.

Everything rides on one thing: write the condition so completion is checkable. The app's own hint says the same: "describe it so you can tell when it's done, e.g. /goal draft is proofread with no typos."

/goal all trip notes are merged into one itinerary with no conflicts

While a goal is active, its row stays pinned to the bottom of the thread, and the slash picker adds "Goal: clear" ("Release the active goal"). Send /goal clear and it answers "Goal cleared."

The "Stop" link is two-step on purpose. It arms to "Stop this?" and asks again, so a stray tap can't wipe out an hour of work. After a hard stop, a "Keep going" link picks it back up.

Steering: redirect a run without ending it

Deep modes mean long runs, and long runs need the occasional course correction. Typing while the agent works won't interrupt it: the composer placeholder switches to a "Responding" hint, Enter to queue, ⌘Enter to send now.

A queued message parks in a strip above the composer, shimmering, with the tooltip "Picked up at the agent's next step". Beside it sit "Send now", "Edit and requeue", and "Remove" buttons. It turns into a chat bubble only once it's actually delivered, marked by a quiet "Delivered your message now" line.

Long shell commands don't hold the thread hostage either. The agent can park one as a background task, shown as a "Running in background" chip with a ticking clock, the command text, and its own stop button ("Stop background command").

When it's done, the chip flips to "Ran in background" and expands to show the tail of the output. If it finishes while the chat is idle, the agent wakes up on its own: a "Background task finished" note appears and it picks the work back up.

Sessions: where the long runs pile up

Deep modes leave long chats behind, and they all land in the left sidebar's "Agent" segment, newest first. "New chat" (⌘⇧C), the "Search chats" field, right-click → "Rename", and "Delete chat" all work exactly as they did in "Chat: talking to your workspace". Long histories stay fast, and nothing falls off the bottom of the list.

Which mode when?