Google Antigravity release notes: CLI 1.1.5 controls models and reasoning effort

Letting every task inherit the same model is convenient until quick file sorting and difficult debugging land in the same work queue. Google Antigravity CLI 1.1.5 makes the choice visible: you can select reasoning effort, pin a stable model slug, and give a subagent its own model tier.

Google Antigravity CLI is the terminal interface for Antigravity, Google's agentic development platform. Reasoning effort controls how much model work a task receives, letting the user trade latency for depth.

Google Antigravity release notes: what changes in CLI 1.1.5?

Version 1.1.5 was published on July 21, 2026. The new /effort command shows and changes the current model's reasoning effort. You can use the picker with the Left and Right keys or enter /effort <level>. The same choice is available at launch through the --effort flag.

The release notes also introduce stable, user-facing model slugs. A model slug is a model's short identifier. It appears in /model, is accepted by --model, and can pin the same model across sessions. The redesigned model picker groups choices by base model and shows an effort badge when a model exposes more than one reasoning variant.

Custom agents also gain a model field in their frontmatter. A subagent can use a tier such as flash or pro, while the default value, inherit, uses the parent agent's model.

Source: Google Antigravity CLI 1.1.5 release notes.

Why model choice per task is useful

An automation team can now route a routine check through a faster model tier and reserve more reasoning for debugging, reconciliation, or architecture decisions. The choice is easier to inspect afterward: the status line shows effort for models with multiple variants, and a stable slug can be recorded in a run recipe.

Start with one focused work queue and record the model slug, reasoning effort, and review point in the run log. You can then compare latency and output quality without confusing a model change with a prompt change. Use only the models and effort levels shown by your installed version.

Short example: use the new Antigravity feature

Human step

Open /model in CLI 1.1.5 and choose from the models shown by your installation. If the model has multiple reasoning variants, use the Left and Right keys in the model picker or open /effort. The release notes also confirm the /effort <level> form, but do not list the level names. Use a value shown locally rather than guessing.

Source for the commands and pickers: Google Antigravity CLI 1.1.5 release notes.

Then copy the agent instruction:

Review the month-end reconciliation workflow.
Split the work into routine checks and tasks that need deep analysis.
Recommend a model tier and reasoning effort for each part, using only options available in this session.
Explain the trade-off and name the final human review point.
Do not edit files.

A good response should:

• Separate routine checks from deep analysis.
• Use model names and effort levels that actually exist in the session.
• Tie each choice to latency or the need for deeper reasoning.
• End with a clear human check before the next action.

Clearer MCP results and background work

CLI 1.1.5 also fixes two problems that show up in integrated workflows. Authenticating a remote MCP server no longer crashes the /mcp panel, and MCP results with embedded resources now surface text and inline media in the conversation instead of dropping them. MCP, the Model Context Protocol, is an open protocol that connects an AI agent to external tools and data sources.

Long-running background work now uses a shared lifecycle with deterministic startup and shutdown. A failure in one background task should no longer disrupt the session, and rapid repeated refreshes are coalesced to reduce redundant work. The release notes also mention a scrollable /settings panel, corrected permission parsing for quoted shell metacharacters, and safer handling of multibyte characters in file search.

Source: Google Antigravity CLI 1.1.5 release notes.

This is a good time to record model choice in the workflow's run recipe. If several agents and MCP tools need to work together, Hammer Automation's Tool Forge can turn model selection, permissions, approvals, and run logs into an operating routine.