Google Antigravity

Google’s agentic coding platform — pitched as an “agent operating system” for building and shipping AI-driven apps across desktop and terminal. It’s Google‘s developer-facing agent IDE/harness, the front end to the same big-tech agent stack ADK serves from the SDK side. Ingested 2026-06-28 from a geeky-gadgets write-up of the 2.0 update (T3 — secondary product-news coverage, not primary docs; treat feature specifics as reported, not confirmed).

What the 2.0 update changes

The headline shift: Antigravity split one unified app into two:

a VS Code-inspired IDE for writing code, and
a separate chat-based agent interface for driving the agents.

The stated goal is specialized workflows per surface; the reported reception is mixed — some users preferred the single unified app, and there are complaints about tighter usage limits after the split. So 2.0 reads less as a capability leap than a repackaging of the same harness into two front ends, with the usual cost-control friction the spoke keeps seeing (anthropic-agent-sdk-billing-pause is the sharper version of the same tension).

What’s in the harness

Mapped onto this spoke’s anatomy, Antigravity carries the now-standard parts:

Parallel sub-agents — “dynamic sub-agents” that run multiple tasks at once (agent-orchestration).
Async / scheduled background work and autonomous operation in remote Linux environments — the long-running/durable axis.
Command-driven control — /goal, /drill me, /browser slash commands; browser/computer use is wired through Chrome DevTools for live debugging.
Programmatic surface — an Antigravity SDK for agent infrastructure plus a CLI with customizable keybindings and themes.
Customizable JSON hooks for integrations — the hook-the-loop pattern surfaced as config, the same shape Claude Code’s hooks (and this hub) run on.
Multi-folder projects with per-folder permissions (scoped autonomy), live voice transcription, and artifact-review flows for collaboration.

A second face — the orchestration engine

The 2026-06-28 ingest paged Antigravity as a developer-facing IDE. A later case study, ai-race-coach (Google Developers Blog), shows a different face: Antigravity as a runtime “domain-bridging engine” that “handled stateful orchestration and telemetry ingestion” for a real-time AI driving coach at Sonoma Raceway. There it isn’t where you write code — it’s the surface that lets engineers work at the level of system behaviour without domain expertise (racing), doing the stateful orchestration under an ADK-driven agent team. So the product spans both the authoring surface (IDE) and a runtime orchestration surface — worth watching whether those are one engine wearing two hats or a genuinely split offering. (Case-study framing, T3 — treat “domain-bridging engine” as the marketing description, not a documented API.)

Availability

No general pricing disclosed in the source; AI Ultra subscribers reportedly get extra credits. Volatile — a product in active, contested iteration.

Where it sits

This is the harness-vendor stratum again: Google now ships both the kit (adk, SDK-side) and a first-party agent IDE (Antigravity, developer-facing), the same dual move Microsoft makes with Foundry + Agent Framework. The split-into-two-apps decision is a live bet on IDE-surface vs chat-surface as separate control points — echoing the same-anatomy-different-control-point frame, but within one vendor’s product line rather than across rivals. The mixed reception (usability regression, tighter limits) is the recurring harness-as-product cost/UX friction, not a technical failure.

google · adk · agentic-coding-harness · agent-orchestration · durable-agents · agent-middleware · ai-race-coach