Exploring the Boundaries of Autonomous Agent Capabilities
A dedicated experimental sandbox designed to test, benchmark, and document Antigravity CLI and Gemini capabilities, customization hierarchies, context injection dynamics, and advanced orchestration patterns.
Rules & GEMINI.md
Investigating contextual guardrails, hierarchical directory inheritance, always-on vs model-decision triggers, 24KB file limits, and the 20,000-token aggregate rules budget.
Actions & Slash Workflows
Cataloging built-in agent actions (/plan, /goal, /schedule, /browser, /boost, /teamwork),
sidecar execution hooks, and asynchronous task control mechanisms.
Context Dynamics & Ingestion
Analyzing progressive skill disclosure (~97% token savings), @-mention entity resolution, line slicing (99% savings),
and directory indexing mechanisms.
Extensibility & MCP Plugins
Prototyping custom Model Context Protocol (MCP) servers, Stdio JSON-RPC 2.0 tool schemas, plugin packaging,
and CLI management commands (agy mcp).
Multi-Agent Swarms & Subagents
Exploring subagent delegation (invoke_subagent), isolated workspace branching via ephemeral Git worktrees,
inter-agent message passing, and automatic cleanup.
Permissions & Security Models
Auditing Antigravity Hub project permissions, wildcard allowlists, tool execution policies, and terminal sandboxing constraints.
📋 Agent Starting Off Point
AGENT_INSTRUCTIONS.md
The project root contains detailed operational directives and deep-dive checklists in
/home/mark/projects/agy/AGENT_INSTRUCTIONS.md and RESEARCH_MISSION.md.
New agent conversations can reference these documents directly to execute targeted experiments.
🛰️ Environment Telemetry
- Subdomain:
agy.85-239-240-162.sslip.io- Internal Port:
127.0.0.1:3302- Service:
agy.service(systemd)- Workspace:
/home/mark/projects/agy