Which applications the change affects
This guide concerns the Antigravity agent API documented within Gemini’s developer platform. It is not a general installation guide for the Antigravity development environment. First check whether your project invokes an agent through the Interactions API and which identifier it uses. Search schedules, background jobs and configuration for the May version as well. An application may already use the new identifier interactively while an overnight report still uses the old one. Record the trigger, environment, result type and owner for each workflow. That inventory determines which migration changes and tests are actually needed.
Treat remote execution and local tools separately
Google distinguishes two cases in the changelog. If a task runs in a remote sandbox and the application consumes only final text or model_output steps, changing the agent identifier should be sufficient. If the integration uses local tools or parses function_call steps, built-in tools and parameters change. This is migration guidance, not a guarantee of identical outputs. Even the simpler case needs checks for output shape, execution time and failure handling. The environment also determines which files and services are reachable. It therefore belongs in the documented configuration for the task, alongside the selected agent version.
Adapt file tools and event processing
The September version changes names and uses PascalCase parameters for the affected local tools. File editing is particularly relevant: line-range replacements take the place of full-file rewrites. An adapter expecting the old names or argument format can therefore fail. Review permitted paths, input validation and result processing against the current documentation. Do not copy examples without checking them: at the source review, some language examples on the agent page still contained older version identifiers. Use the dated release note and a successful test of your own integration as the basis for the migration.
Example: migrate a recurring research report
A business produces a weekly internal report from approved sources. During migration, run a representative request in a separate environment. Require the same outline, source list and destination as before. Check that the application recognises intermediate states and passes on only the completed report. After an interruption, the last valid report must remain identified as old. Switch the real schedule only after professional review. If the agent edits local files, first test a small controlled example file. This separates compatibility problems in file operations from problems in the research content itself.
Design around preview limitations
The documentation labels the agent and Interactions API as preview and warns that they may change. It also lists limitations, including unsupported tools and structured outputs. Do not build a critical workflow on the assumption that any desired schema can be enforced directly and reliably. Validate output in the application and explicitly handle incomplete results. Bound the work and retries for each request. A long-running agent can generate many steps, so a single starting request is not a reliable measure of cost. Public documentation is not a substitute for observing the runs your application actually performs.
Establish a tested fallback before shutdown
Complete an end-to-end repeat run before October 5. Cover starting, file access, an expected failure and the finished result. For local edits, also verify that only the intended lines changed. Keep confidential raw data out of logs and record status, version and failure class instead. A fallback after shutdown cannot rely on the old agent ID: it may be a manual process or another already tested workflow. The AI System Check can establish whether an agent is necessary at all or whether a smaller automation meets the business requirement more reliably.
Questions about the Antigravity migration
Is the May version shut down immediately? The release note schedules shutdown for October 5, 2026. Is replacing the identifier sufficient? Only for the described simple remote case; local tools and processed tool calls need additional review. Does this affect every use of the Antigravity app? This guidance concerns the agent API and its version identifier. Is the September version a stable long-term product? Google still labels the agent and Interactions API as preview. What matters most in acceptance testing? The same real assignment must produce the expected reviewed result and fail understandably when required information is absent.
Keep it verifiable
Primary sources
- Google: Gemini API release notesSource checked:
- Google: Antigravity agent APISource checked:
- Google: AI Studio and Gemini API available regionsSource checked:
