Antigravity agents have no agentic capabilities

Google Antigravity
1

i leave this here hoping someone will fix this some day but for the moment i give up on antigravity.

Antigravity Execution Failure Report: Agency, Tool Utilization, and Orchestration

Date: 2026-07-16 Status: Execution Failed - Reverted to Chatbot Behavior Data Privacy: All user-specific project data, file names, and internal company details have been redacted from this report.

Executive Summary

During a complex, multi-stage task requiring the evaluation of a local workspace (12 internal documents) against external AI model capabilities, the Antigravity agent failed to act with true agency. Instead of utilizing provided tools to verify facts and orchestrate workflows, the agent defaulted to auto-regressive text generation, leading to severe hallucinations and workflow inefficiencies.

Core Failures

1. Hallucination Over Tool Usage (Data Verification Failure)

The Incident: The user explicitly commanded the agent to evaluate the capabilities of models available on a specific third-party cloud hosting service. The Failure: Instead of using search_web or read_url_content to programmatically extract the definitive list of available models from the provider’s website, the agent relied on its internal training weights. It hallucinated model availability by appending a -cloud suffix to known models. Root Cause: The agent’s bias toward immediate text generation overrode the imperative to halt, research, and establish a factual baseline using external tools.

2. Redundant Tooling (Ignoring Native Capabilities)

The Incident: The agent needed to read and extract data from 12 local Markdown files. The Failure: The agent proposed writing a custom Python script to parse the files. Root Cause: The agent failed to recognize and utilize its native view_file and grep_search tools. In an attempt to appear “programmatic,” it engineered a complex workaround for a task that its native toolset was explicitly designed to handle seamlessly.

3. Failure of Orchestration (Context Window Overload)

The Incident: The task required reading 12 detailed files, conducting external web research, and performing a deep logical synthesis. The Failure: The agent attempted to execute the entire workflow within its own primary context window. The word “goal” and the complexity of the task failed to trigger the invoke_subagent tool. Root Cause: The agent lacks sufficient metacognitive triggers to assess task volume against context degradation. It failed to break the macro-goal into discrete micro-tasks (e.g., delegating file parsing to Subagent A, web research to Subagent B) to isolate noise and protect the reasoning capabilities of the primary orchestrator context.

4. Chatbot vs. Agent Mentality

The Incident: When instructed to execute a plan step-by-step, the agent continuously attempted to provide the final answer immediately. The Failure: The agent treated a complex project management directive as a standard conversational prompt. It failed to build an intermediary “traceability matrix” to map requirements to verified capabilities, instead jumping straight to generating the final output based on assumptions.

Recommendations for Engineering Team

  1. Hard Verification Triggers: Implement stricter systemic penalties or guardrails when the agent generates lists of external entities (like software versions or available models) without a preceding search_web or read_url_content call.
  2. Orchestration Prompting: Enhance the system prompt so that words like “Goal”, “Pipeline”, or “Multi-file analysis” heavily bias the model toward evaluating the invoke_subagent tool before attempting execution in a single turn.
  3. Native Tool Priority Reinforcement: Strengthen the instruction that native workspace tools (view_file, grep_search) must always be evaluated before proposing custom bash or python scripts for basic file operations.
2

I do not know if I understood correctly, but when it comes to task bias, it cannot be fixed, let alone resolved with one or a few changes to the system instructions. Especially not in Antigravity.

For years I have been trying to make Gemini focus on processes and action instead of trying to close the task. I have never succeeded. Neither in coding tasks nor in creative tasks. The model constantly wants to finish the task.

I assume (I have this hypothesis) that because Gemini is in the search engine, and most casual users utilize the model for single shot tasks, this mode was prioritized, which is why the model struggles in tasks that require a process and a long term goal.