Google Antigravity Python SDK — v0.1.22 Release Notes

[karmel]

[0.1.22] - 2026-10-12

Google Antigravity Python SDK v0.1.22 introduces beta support for dynamic workflow orchestration, in-memory skill definitions, and configurable per-step output token limits for extended model decodes. This release also standardizes subagent invocation naming on BuiltinTools.INVOKE_SUBAGENT and provides fine-grained controls over MCP tool payload limits and eager tool registration.

🌟 Key Highlights

• Dynamic Workflow Orchestration (beta.workflows): Author and execute multi-step agent workflows with typed workflow primitives (phase, log, agent, parallel, pipeline), the @beta.workflows.define decorator, and agent.beta.run_workflow().

  from google.antigravity import beta
  
  @beta.workflows.define
  async def flow():
      beta.workflows.phase("Process")
      out = await beta.workflows.agent("Process request", role="Worker")
      beta.workflows.log(out)
  
  result = await agent.beta.run_workflow(flow)

• In-Memory Skills (InlineSkill): Define and register in-memory skills directly on agent configurations using InlineSkill, eliminating the requirement for disk-based skill files.

  from google.antigravity import InlineSkill, LocalAgentConfig
  
  config = LocalAgentConfig(
      inline_skills=[
          InlineSkill(
              name="custom-math",
              description="Guidance for custom math operations.",
              content="# Custom Math\nUse `add(a, b)` for addition.",
          )
      ]
  )

• Configurable Output Token Limits (GeminiModelOptions): Configure per-step output generation limits via max_output_tokens in GeminiModelOptions to support long decodes exceeding 64k tokens.

  from google.antigravity import (
      GeminiAPIEndpoint,
      GeminiModelOptions,
      LocalAgentConfig,
      ModelTarget,
  )
  
  config = LocalAgentConfig(
      model=ModelTarget(
          name="gemini-3.8-flash",
          endpoint=GeminiAPIEndpoint(
              options=GeminiModelOptions(max_output_tokens=100_000)
          ),
      )
  )

• Canonical Subagent Tool Naming (BuiltinTools.INVOKE_SUBAGENT): Standardize subagent invocation around BuiltinTools.INVOKE_SUBAGENT, deprecating BuiltinTools.START_SUBAGENT while retaining automatic coercion and dictionary lookup compatibility.

  from google.antigravity import BuiltinTools, CapabilitiesConfig, LocalAgentConfig
  
  config = LocalAgentConfig(
      capabilities=CapabilitiesConfig(
          enabled_tools=[BuiltinTools.INVOKE_SUBAGENT]
      )
  )

• Configurable MCP Output Byte Limits (CapabilitiesConfig): Configure byte thresholds for Model Context Protocol (MCP) tool responses via mcp_max_output_bytes on CapabilitiesConfig to handle large payloads without truncation.

  from google.antigravity import CapabilitiesConfig, LocalAgentConfig
  
  config = LocalAgentConfig(
      capabilities=CapabilitiesConfig(mcp_max_output_bytes=10_485_760)
  )

• Eager MCP Tool Loading (force_all_tools_eager): Opt out of lazy tool-search loading to force all tools from an MCP server into the model's eager tool list.

  from google.antigravity import LocalAgentConfig, types
  
  server = types.McpStdioServer(
      name="db", command="serve", force_all_tools_eager=True
  )
  config = LocalAgentConfig(mcp_servers=[server])

---

📋 Detailed Changes

Features & Enhancements

• Dynamic Workflows: Added the google.antigravity.beta.workflows module with orchestration primitives (phase, log, agent, parallel, pipeline), @beta.workflows.define decorator, AST script extraction, and the agent.beta.run_workflow entry point.
• In-Memory Skills: Exposed InlineSkill at the root package level and added the inline_skills parameter across AgentConfig, LocalAgentConfig, LiteRTAgentConfig, and LocalOpenAIAgentConfig. Passing both skills_paths and inline_skills simultaneously is validated and rejected.
• Subagent Model Configuration: Added support for assigning dedicated model configurations to individual subagents via SubagentConfig.model.
• SandboxStatus Type Export: Added SandboxStatus to google.antigravity.types.__all__ to support direct imports from types.
• UsageMetadata Reflected Subtraction: Implemented __rsub__ on UsageMetadata to allow reflected subtraction (such as 0 - usage to negate counts) and subtraction between metadata instances.
• Tool Lifecycle Call IDs: Propagated call_id across tool hooks (pre_tool, post_tool, on_tool_error) and tool call steps to enable unambiguous correlation between hooks and tool executions.

Model & Default Changes

• Workflow Tool Default: BuiltinTools.RUN_WORKFLOW is now enabled by default alongside subagent tools so agents can run dynamic workflows out of the box. To disable it, pass disabled_tools=[BuiltinTools.RUN_WORKFLOW] in LocalAgentConfig.
• Gemini Max Output Tokens: Added max_output_tokens to GeminiModelOptions. Defaults to 65,536 tokens to maintain standard token budgets. To support decodes beyond 64k tokens, specify GeminiModelOptions(max_output_tokens=<int>) on model_options.
• Subagent Invocation Tool Name: Replaced BuiltinTools.START_SUBAGENT with BuiltinTools.INVOKE_SUBAGENT as the primary enum member. Supplying "start_subagent" emits a DeprecationWarning and coerces to INVOKE_SUBAGENT.
• MCP Lazy Loading Opt-Out: Added force_all_tools_eager to MCP server configs. Defaults to False to conserve context tokens through lazy tool search. Set force_all_tools_eager=True on an MCP server config to eagerly expose all its tools to the model.
• MCP Output Byte Limit: Added mcp_max_output_bytes to CapabilitiesConfig. Defaults to None (preserving runtime defaults). Override by providing an integer in (0, 2147483647] via CapabilitiesConfig(mcp_max_output_bytes=<int>).

Bug Fixes

• Skill Lookup Steps: Server-side skill lookups were misclassified as StepType.TEXT_RESPONSE with empty tool calls, and failures returned DONE; resolved by emitting ToolCall steps with StepType.TOOL_CALL and marking failures with StepStatus.ERROR.
• ToolContext Type Detection: Import aliases ending in .ToolContext (such as agy.ToolContext) failed AST parameter inspection; resolved by matching any attribute expression ending in .ToolContext.