agent-estate

Agent Estate

Perpetual autonomous work loop plugin for Claude Code. Turns Claude into an agent that works cycle after cycle, maintaining a persistent ledger across all sessions, never stopping until explicitly told to.

Overview

Agent Estate creates an infinite work loop by intercepting Claude's stop events via a shell hook. Each time Claude tries to exit, the hook blocks the exit, increments the cycle counter, and re-injects the prompt — creating a true perpetual agent.

Architecture

agent-estate/
├── .claude-plugin/plugin.json   # Plugin manifest (name, version, author)
├── commands/
│   ├── start.md                 # /agent-estate:start — activates the loop
│   ├── stop.md                  # /agent-estate:stop — removes state file to allow exit
│   └── status.md                # /agent-estate:status — shows cycle count, ledger summary
├── hooks/
│   ├── hooks.json               # Registers stop-hook.sh on the "Stop" event
│   └── stop-hook.sh             # Core engine: blocks exit, increments cycle, re-injects prompt
├── scripts/
│   └── setup-estate.sh          # Creates .claude/agent-estate.local.md state file
├── LICENSE
└── README.md

Key Components

Stop Hook (hooks/stop-hook.sh)

The core engine. Runs on every Claude "Stop" event:

1. Checks if .claude/agent-estate.local.md exists (is the loop active?)
2. If not active: exits normally (exit 0)
3. If done (done: true in frontmatter): removes state file, exits normally (auto-stop)
4. If active and not done:
   • Detects rate limits (429) → waits 60s
   • Detects API overload (529) → waits 30s
   • Increments cycle counter in the state file
   • Outputs a JSON {"decision": "block", "reason": "..."} to prevent exit
   • Re-injects the prompt for the next cycle

Setup Script (scripts/setup-estate.sh)

Creates the state file at .claude/agent-estate.local.md with YAML frontmatter (active, cycle, started_at, user_prompt) and the full cycle protocol instructions. If already active, outputs current state and continues.