The April 2026 Antigravity release is what I'd call a quietly important one. There's no single headline-grabbing launch, but there are seven changes that meaningfully improve the day-to-day for anyone using Antigravity in real work.
Google's release notes are split per feature, which makes it hard to know what to look at first. This roundup re-orders the same features by the order I'd actually pick them up if I had only an hour to play with the update.
If you're short on time, just read sections 1–3 (Gemma 4 + AgentKit 2.0). Those cover roughly 70% of the practical impact of this release.
1. Gemma 4 integration is no longer beta
The most underreported change of the month. The Gemma 4 integration that's been in beta since early 2026 has been promoted to a fully supported feature. The headline doesn't sound dramatic, but the operational consequences are.
While it was in beta, there was no SLA, which is why I avoided pointing production traffic at Gemma 4 via Antigravity. Going GA brings documented error rates, an explicit local-LLM fallback contract, and clear region-failover behavior — everything you actually need to ship something serious.
In the Models tab, picking Gemma 4 and enabling use_local_fallback: true will now seamlessly switch to a local Gemma 4 instance during transient Google Cloud incidents. I run this on a MacBook Pro M4 alongside the cloud setup; quality degradation during fallback is barely perceptible in practice.
2. AgentKit 2.0 ships
This is the headline feature of April. Three changes from AgentKit 1.x stand out:
First, official multi-agent orchestration primitives. The "parent agent delegates to child agents" pattern that you used to write by hand now has first-class support. This alone removes a meaningful chunk of boilerplate.
Second, the A2A (Agent-to-Agent) protocol implementation. Multiple agents can now talk over a shared protocol, which let me wire up a "code review → run tests → deploy" chain through GitHub Actions in under a day.
Third, the new Inspector view. You can now step through an agent's reasoning and see exactly why it picked a particular tool call. This was the single biggest pain point of AgentKit 1.x for me, and the personal highlight of the entire April release.
# Minimal multi-agent example with AgentKit 2.0
from antigravity.agentkit import Agent, Orchestrator
researcher = Agent(name="researcher", role="Find relevant docs", model="gemma-4")
writer = Agent(name="writer", role="Synthesize findings", model="claude-sonnet-4-6")
orchestrator = Orchestrator(agents=[researcher, writer])
result = orchestrator.run("Summarize 2026 AI agent trends")
# Inspector view auto-records each agent's decisions
print(result.inspector_url)3. Project Context retention extended
A small change with a real effect on my flow: Project Context retention went from 7 days to 30 days.
Before, opening a project on Monday after a long weekend often meant waiting for context to re-load. With 30 days, the agent now reliably remembers prior decisions across normal work patterns. The continuity feels different in a way that surprised me.
The catch: longer retention means cross-project context contamination is now a more real risk. If you're switching to a fundamentally different project, get into the habit of running Settings → Project Context → Clear deliberately before you start.
4. Unreal Engine 5 support is GA
Slightly outside my own day job, but worth flagging: UE5 support is now official. You can edit .uasset files directly from Antigravity, and Blueprint generation is wired in.
A friend who works in game dev tested it for me. His take: simple Blueprint generation accuracy is around 70%, but the explanation of UE5 docs is genuinely impressive — good enough to be useful as a teaching tool right now, even if you wouldn't trust it for production logic yet.
I wrote a separate deep-dive in Antigravity × Unreal Engine 5 Integration Guide for anyone working in UE5.
5. Localization improvements
If you'd been bothered by the Japanese UI translation, the April revision brings the agent-related terminology in line with industry conventions ("orchestration", "handoff", "tool calling"). For internal docs and team enablement materials, this saves a real amount of friction.
6. The 168-hour lockout issue is largely fixed
The infamous "Antigravity locked my account for 168 hours" complaint that surfaced in late March has been addressed. The detection logic for failed login attempts has been reworked, and legitimate users mistyping their password a few times are now far less likely to get locked out. One of my teammates was burned by this in March; we haven't seen it since the April update.
If you're still locked out from before the fix, the recovery procedure is in Antigravity 168-Hour Lockout Recovery Guide.
7. Real-time quota visibility
Quietly nice: the sidebar now shows real-time quota consumption at all times. Previously it was easy to discover you'd hit your monthly cap by surprise. Having remaining-call counts always visible at the bottom-right makes pacing your work near month-end straightforward.
The quota algorithm itself didn't change, but the visibility is genuinely felt.
What I'd bet on for May–June
These are personal predictions, not announcements:
- Gemma 5 support, somewhere mid-to-late May, based on Google's typical release cadence.
- Native Claude Opus 4.6 / Sonnet 4.6 integration with one-click switching from the model picker (today you set it via API key).
- Mobile apps, with iOS/Android betas likely around Google I/O (late May).
If you only try one thing from this update, open the AgentKit 2.0 Inspector view on a real agent task. The "why did it do that" visibility alone changes how it feels to build with agents.