It thinks out loud
Instead of a bare answer, the model surfaces its constraints, assumptions, proposals and insights, so you can steer the reasoning, not just react to the result.
Back to the extensions
Spaces extension · the flagship Keep ten fronts
Summary 11m 00s ago Direction Questions 11m 00s ago 
Keep ten fronts
moving at once.
The workflow the whole product is built around: run many spaces in parallel, spawn visible, steerable subagents, and get returns bounded by information, not by turns. You answer the front that needs you; the rest keep working.
Check-inTranscriptScheduledCommitsSettingsAwaiting input·11m 00s
Assessment
The split into ingest / api / worker is sound. One sharp edge: the worker and the api both write sessions, so the boundary has to own a single writer — or you reintroduce the exact race we just fixed in the token store.
Trade-offs
- A dedicated session service is cleanest, but adds a hop.
- Folding writes into the api keeps the topology smaller.
Answer below, or type a direction in plain language.
1## Proposals
2
3### Who owns session writes after the split?
4
8
9Recommendation: api owns the write, the worker calls it.
10One writer, no new service, the boundary stays honest.
Click a tab to switch fronts; answer the one that’s hit an information boundary while the rest keep working.
How it works · the Spaces extension 01 02 03 04
The bottleneck is you now. Spaces raises the signal.
Agents produce faster than anyone can review: walls of diffs, walls of text. So Spaces makes every return count. The model keeps working until it hits a genuine information boundary, a decision only you can make, and when it comes back, it's done the thinking, not just dumped output. To get there, it makes each exchange far richer than a chat message.
It holds a living context
A running picture of the work stays pinned in view. Focus survives even on long sessions, and you can drop back in after an hour without re-reading everything.
It draws
When words run out, it sketches architecture, flows, and trade-offs on an Excalidraw canvas you can edit and hand straight back.
You reply your way
Answer a structured questionnaire, write freeform direction, or draw your response. Whatever closes the gap fastest.
The result: fewer, higher-quality returns and longer productive runs. The model only taps you on the shoulder when it genuinely needs a decision, so you can hold more fronts at once without drowning in half-finished output.
Subagents
Subagents you can actually see.
In most tools a subagent is a black box: it spawns, it churns, and you learn how it went only when it returns, too late to stop the tokens it just burned going the wrong way. Frontier's subagents are spaces nested under spaces: out-of-process, fully visible, and yours to step into mid-run.
Watch any of them, live
Open a child and you're inside its work: its reasoning, its proposals, its diffs. No spinner, no waiting to find out how it went.
Step in mid-run
Answer a question, correct a wrong turn, or take the wheel, all without killing the work it's already done. They're steerable, not fire-and-forget.
Each on its own model
A child can run Claude while the parent runs GPT, or a cheap local model for the grunt work. The right tool on every branch of the tree.
It’s just an extension
Spaces is a flagship core extension. Retune it, fork it, delete it, or write your own.
Download See the rest of the extensions →
Frontier’s desktop app runs on macOS, Windows & Linux — open this page on your computer to download.
macOS Windows Linux