Antigravity Boost Deep Reasoning: How /boost Works

Editorial illustration of /boost deep reasoning in Antigravity: focused workstream agents with independent verification badges converging on a task. — /boost is a per-task escalation, not a default. Generated hero; not official artwork.

What /boost Is

The launch post describes /boost precisely: the Antigravity harness can tackle a wide range of tasks right out of the box, but /boost helps if you are tackling particularly complex tasks and are willing to spend some more. The “some more” is the honest part — boost is a spend-more-to-verify-more feature, not a free quality bump.

Introducing /boost in Google Antigravity. The Antigravity harness can tackle a wide range of tasks right out of the box, but /boost helps if you are tackling particularly complex tasks and are willing to spend some more.
— @antigravity August 31, 2026

Mechanically: when you invoke /boost, the Antigravity harness delegates the task to focused workstreams, and the results pass through independent verification across iterative rounds before anything reaches your working tree. The verification layer is the differentiator — the model checks its own work with fresh context instead of trusting its first pass.

How It Works

The launch post does not publish the internal mechanics — how many workstreams, how verification is implemented, which models verify. What the docs do document is the surface behavior: invoke with /boost in the agent panel, describe the task, and the harness runs the deeper pass. The verification round structure means slower wall-clock time per task in exchange for higher confidence in the result.

What you can verify yourself: run the same task with and without /boost on a problem you know the answer to, and compare. That is the only benchmark that maps to your codebase, and it directly measures the spend-versus-confidence trade the feature is built on.

When It Is Worth It

Boost is worth the spend where a wrong answer costs more than the extra tokens. The task classes that fit:

Cross-file refactors where a subtle miss breaks a build far from the change.
Concurrency and race-condition work where the bug only reproduces under specific interleavings.
Security-adjacent changes — auth flows, permission checks, data validation — where verification rounds catch what a single pass would confidently miss.
Large merges and migrations where the diff spans enough files that human review alone cannot cover the blast radius.

The flip side: boost is waste on mechanical work. Renames, doc updates, and simple bug fixes gain nothing from verification rounds — you pay multiplied tokens for a second opinion on a task that did not need one.

/boost vs High Thinking

These are different tools that look similar from the outside. High thinking is one model reasoning deeply in a single pass — the cost is linear in reasoning tokens. Boost delegates the harness to multiple focused workstreams with iterative verification — the cost is multiplied by the workstream count and the verification rounds.

Approach	Mechanics	Spend	Best for
Single run, High thinking	One model reasoning deeply in one pass	Feature + output tokens	Hard single-file problems, deep single-domain reasoning
/boost	The harness delegated to focused workstreams, iterative rounds, independent verification	Multiplied: each workstream draws its own tokens	Multi-workstream problems where one pass is not enough

The decision rule that falls out: use High thinking when the problem is hard but bounded to one context. Use boost when the problem spans contexts or the cost of a wrong answer is high enough to justify verification. Neither is a default — both are per-task escalations.

Official Demo

The design-system demo from the 3.8 Flash launch is the closest official illustration of the workload class boost targets: large-context generation where independent verification of a broad output matters:

Official Google Antigravity demo: the broad-output workload class where verification matters.

Caveats

Verification is not proof. The verification rounds are model-driven, not test-suite-driven — a model checking its own work with fresh context catches different errors than your test suite does. Keep both.

No published cost multiplier. The launch post does not document how many workstreams boost runs or how the token spend scales. Your own before/after measurement on a known task is the only number that is about your quota.

Not a substitute for human review on risky changes. The verified output still lands in your working tree for your review. Boost raises confidence; it does not transfer responsibility.

Verdict

Use /boost for the task classes where verification rounds earn their cost: cross-file refactors, concurrency work, security-adjacent changes, and large migrations. It is the right escalation for exactly those and the wrong default for everything else.

Skip it for mechanical work — renames, docs, simple fixes gain nothing from multiplied tokens. And benchmark it against your own tasks before wiring it into any pipeline, because the spend is real and the verification is model-driven, not test-driven.

FAQ

What is /boost in Antigravity?

Antigravity’s deep-reasoning escalation. The harness delegates the task to focused workstreams with independent verification across iterative rounds — for complex tasks where you are willing to spend more.

How is /boost different from High thinking?

High thinking is one model reasoning deeply in a single pass; the cost is linear. /boost delegates to multiple workstreams with iterative verification; the cost is multiplied. Use High for hard-but-bounded problems, boost for problems that span contexts.

How much does /boost cost?

Google does not publish a cost multiplier. It draws more tokens than a standard run — multiplied by workstreams and verification rounds. Run the same task with and without boost on a known problem to measure your own spend.

How do I invoke /boost?

Run /boost in the Antigravity agent panel and describe the task. The deeper pass runs automatically; verified results land in your working tree for your review.

Is /boost verification reliable?

It is model-driven, not test-suite-driven — it catches different errors than your tests. Keep both. And the verified output still lands in your working tree for your review; boost raises confidence but does not transfer responsibility.

Should /boost be my default?

No. It is a per-task escalation for high-cost-of-error work. As a default, it multiplies token spend on mechanical tasks that gain nothing from verification rounds.

Sources

Official sources first. All accessed September 2–9, 2026.