Google Antigravity adds /boost multi-agent deep reasoning for hard coding tasks

Google ships a deep-reasoning mode for Antigravity

Google has added a new slash command to Antigravity, its agentic coding environment. Typing /boost activates an on-demand multi-agent reasoning pipeline built for hard software engineering problems, the kind where ordinary single-turn assistance tends to stall. According to the Antigravity documentation, which surfaced on Hacker News' front page, the feature targets the gap between fast interactive coding and fully autonomous multi-day agent campaigns.

What /boost is for

The documentation splits development work into three difficulty bands. Everyday engineering, such as feature work, codebase navigation and routine refactoring, is handled by the standard agent. Long-horizon campaigns, like repository-scale migrations and multi-day exploratory research, belong to the separate /teamwork-preview mode. /boost covers the middle ground: high-intensity interactive tasks such as concurrency bugs, algorithmic optimisation, subtle regressions and intricate multi-file refactoring. The docs emphasise that it requires no complex setup or prolonged scoping interview; you invoke it and execution starts immediately.

How the three-phase pipeline works

/boost runs a process that deliberately separates strategy from execution and verification.

In phase one, goal and strategy formulation, a Primary Orchestrator receives the prompt, inspects workspace context, and decomposes the challenge into discrete, verifiable subtasks, deciding which specialist workstreams are required.

In phase two, parallel execution and verification, the Orchestrator dispatches those subtasks to specialised subagents operating in clean, isolated scopes. Implementation workstreams construct candidate solutions, apply refactors and generate unit tests. Investigation workstreams perform root-cause debugging, trace execution call graphs and analyse unfamiliar dependencies without modifying files. A local verification lane executes build targets and test suites to validate hypotheses before results are reported back.

In phase three, synthesis and delivery, the Orchestrator aggregates the findings and validates the combined solution against full test suites and edge cases. If an assertion fails, the error diagnostics feed into the next iteration for automated correction. Only once tests and requirements pass does it deliver a concise summary with the verified changes.

Where it sits in Antigravity's lineup

The documentation contrasts three execution modes. The default agent is a single-agent direct loop with a shared working tree and single-pass tool checks; it suits tasks lasting seconds to minutes and is available on all plans. /boost spans seconds to hours, uses a three-phase reasoning hierarchy, ephemeral isolated worktrees and multi-round independent verification, and requires a paid plan. /teamwork-preview handles hours-to-days work with multi-role agent teams, persistent isolated worktrees per milestone and adversarial falsification, but only after a two-phase scoping interview.

Using it, and the guardrails

The command works across Antigravity surfaces: type it at the start of a prompt in Antigravity 2.0, or enter it directly in the CLI's terminal prompt. The documentation's example prompts include reproducing and fixing an intermittent deadlock in a connection pool, implementing a lock-free ring buffer for streaming telemetry with stress tests, and tracing why HTTP request timeouts spike once batch payloads exceed 2MB, the last being an investigation-only task that modifies no code.

On safety, Google says subagents inherit the file-access and command-permission policies configured for the active workspace, and that when a worker proposes a protected terminal command or an edit outside trusted scopes, an authorisation prompt surfaces in the developer's interface. Subagents also run in isolated memory spaces, so verbose debug logs and scratch diffs do not clutter the primary chat history.

Why it matters

The notable bet here is verification, not just parallelism. Multi-agent coding tools have a known failure mode: confident, plausible output that has not actually been tested. By productising a loop of isolated execution, local test runs, regression checks and automated correction, /boost pushes verification discipline into ordinary coding sessions rather than reserving it for heavyweight autonomous runs. The tiering is also telling: deep reasoning sits behind a paid plan while the basic loop stays on all plans, and the three modes form a clear escalation path from quick fix, to hard problem, to multi-day campaign. For developers wrestling with race conditions and gnarly refactors, it offers a middle option that is heavier than a chat completion but far lighter than scoping out a full agent campaign.