Google Antigravity release notes: CLI 1.1.3 searches code and SDK 0.1.7 isolates the runtime

Antigravity received two updates that show up in real workflows. CLI 1.1.3 can search code across an entire workspace. SDK 0.1.7 gives each agent instance its own environment variables and lets local models use MCP servers and subagents. Together, they shorten the path from "where is this integration?" to a run you can review.

Google Antigravity is Google's platform for agentic development. An agentic development platform lets an AI agent read a project, use tools, and complete multistep work in the same workspace. Antigravity CLI is the terminal interface, while Antigravity SDK is the Python library for custom agent workflows.

Google Antigravity release notes: CLI 1.1.3 searches the whole workspace

Antigravity CLI 1.1.3 was released on July 16, 2026, at 01:01 UTC. The new /codesearch command has /cs and /search aliases. Queries are regular expressions by default. The -F or --literal flag switches to exact text matching, while f: or file: uses path globs to include or exclude files.

This is more useful than another general chat search. You can find every place that writes a customer field, calls an API client, or uses an old permission rule before the agent proposes a change.

The release also marks earlier context-compaction boundaries and loads skills, rules, agents, and hooks faster. In print mode (-p), tools that require confirmation should now be soft-denied instead of hanging or being silently approved. The error names the allow rule needed to proceed. Writes outside the workspace should no longer be auto-approved in always-proceed mode, and MCP calls now have timeouts so an unresponsive server cannot hold the agent indefinitely.

Human step: open the correct workspace in Antigravity CLI 1.1.3 and use /codesearch. The release notes verify the command, aliases, and flags above but do not give a complete command example. Follow the interactive help in your installation rather than assuming an argument order.

Source: Antigravity CLI 1.1.3 release notes

SDK 0.1.7 gives each agent a clearer runtime environment

Antigravity SDK 0.1.7 is dated July 14 in the SDK changelog and was published on PyPI on July 16. A runtime environment is the set of variables, tools, and connections a process receives when it starts. In 0.1.7, LocalAgentConfig can receive its own env dictionary. The SDK passes it to the isolated agent process without requiring you to alter the parent process's global environment.

For a real integration, the application can retrieve a scoped API key from a secret manager and pass that key to the agent instance. Pair it with tool policies and an approval point for writes. The agent can then work against a real system with the access it needs, while the run log records which tools it used.

The SDK also adds atomic updates and locking for shared state in multithreaded tools and hooks. Local Gemma models through LiteRT and OpenAI-compatible local endpoints such as Ollama or LM Studio can now be configured with subagents and MCP servers. MCP, or Model Context Protocol, is a standard way to connect an agent to external tools and data sources. Other changes include an extra_high thinking level, more reliable local socket discovery in containers, and a fix that stops the background spinner from overwriting approval questions in the interactive console.

Human step: install the SDK with the officially documented pip install google-antigravity command, or update the dependency through your project's normal package workflow. Load secrets into the runtime and build LocalAgentConfig(env=...) from the values the agent actually needs. Do not hardcode live keys in source files.

Sources: Google Antigravity SDK 0.1.7 changelog, google-antigravity 0.1.7 on PyPI

Antigravity 2.3.1 fixes a startup blocker

The same coverage window contains a narrower app patch. Antigravity 2.3.1, dated July 16, fixes a bug where an empty or malformed config.json file in ~/.gemini/config/ prevented AGY from loading. Google rolls out app versions gradually, so the fix may take a few days to reach everyone.

Source: Google Antigravity 2.3.1 changelog

What Nordic teams should test now

Choose a workflow where several code paths can write to the same business field, such as customer status, stock level, or case status. Use code search to locate every write. Then have the agent map validation, test coverage, and external API calls before any file changes.

If the check should become recurring, SDK 0.1.7 can run it with a dedicated environment, a scoped key, defined tools, and a review step before writes. That is a natural Tool Forge job: turn a one-off review into a clear agent workflow with permissions, stop points, and a run log.

Short example: use the new Antigravity feature

Human step: first search for writes to the relevant business field with /codesearch in CLI 1.1.3. Once you have the search results, give the agent this instruction:

Review the code-search matches for writes to customer_status.
Map each write to its validation, tests, and external API calls.
List missing checks and propose one bounded implementation plan.
Do not edit files.

Good output should include:

• All relevant matches grouped by code path
• Existing validation and test coverage for each write
• Missing checks with file references
• A plan that can be reviewed before the agent receives write access

What to watch next

CLI 1.1.3 documents the search command and filters but not the full interactive query syntax. The next practical check is how it behaves in large monorepos and remote terminals. For the SDK, measure how shared state, local MCP tools, and approval points behave under concurrent runs, not just in a simple demo.