Scheduled Autonomous Agents
Run Claude Code on autopilot. Create agents that execute any prompt on a schedule — morning briefings, overnight builds, recurring research — delivered straight to your Beeper.
Claude Code isn't just interactive. It's a daemon.
HOW IT WORKS
┌─────────────────┐ ┌──────────────┐ ┌─────────────────┐
│ /herobrine │────▶│ macOS │────▶│ claude -p │
│ create │ │ launchd │ │ (headless) │
└─────────────────┘ └──────────────┘ └────────┬────────┘
│
┌────────▼────────┐
│ Deliver via │
│ Beeper + notif │
└─────────────────┘- You define an agent: a name, a prompt, and a schedule
- A macOS launchd plist is created to trigger it on time
- The runner executes
claude -pwith--dangerously-skip-permissions(fully autonomous) - Results are sent to your Beeper chat and saved to local logs
- macOS notification confirms completion
SETUP
1. Install the scripts
Copy the runner and manager scripts to ~/.claude/agents/:
mkdir -p ~/.claude/agents/schedules ~/.claude/agents/logs
cp scripts/run-agent.sh ~/.claude/agents/run-agent.sh
cp scripts/manage-agent.sh ~/.claude/agents/manage-agent.sh
chmod +x ~/.claude/agents/run-agent.sh ~/.claude/agents/manage-agent.sh2. Configure Beeper delivery (optional)
Find your Beeper "Note to Self" chat ID and update the default DELIVERY_CHAT in manage-agent.sh. The agent will send full results as a formatted Beeper message after each run.
If you don't use Beeper, results are still saved to ~/.claude/agents/logs/ and delivered via macOS notification.
ACTIONS
Determine the action
If $ARGUMENTS contains an action keyword (create, list, remove, run-now, logs), use that. Otherwise, ask the user with AskUserQuestion:
- Create a new scheduled agent
- List all scheduled agents
- Remove a scheduled agent
- Run now — trigger an agent immediately
- View logs of past agent runs
Action: Create
Gather from the user (via AskUserQuestion or arguments):
- Agent name — short identifier, lowercase with hyphens (e.g.,
stock-news,nightly-review) - Prompt — the full prompt sent to Claude. Be detailed. Can include skill invocations like
/steveor/gsd:execute-plan - Schedule — friendly options:
- Every morning at 8am → 0 8 * * * - Every evening at 6pm → 0 18 * * * - Every hour → every-1h - Every 30 minutes → every-30m - Weekday mornings at 9am → 0 9 * * 1-5 - Custom cron or interval
- Model —
opus,sonnet(default), orhaiku - Max turns — default:
50
Then run:
bash ~/.claude/agents/manage-agent.sh create "<name>" "<prompt>" "<schedule>" "$HOME" "<model>" <max_turns>Action: List
bash ~/.claude/agents/manage-agent.sh listAction: Remove
If no agent name given, list first, then ask which to remove.
bash ~/.claude/agents/manage-agent.sh remove "<name>"Action: Run Now
bash ~/.claude/agents/manage-agent.sh run-now "<name>"Action: Logs
bash ~/.claude/agents/manage-agent.sh logs [agent-name]USE CASES
| Agent | Prompt | Schedule |
|---|---|---|
stock-news | Research today's top stock market movements and tech earnings. Summarize the 5 most important things. | 0 8 * * 1-5 |
pr-digest | Review all open PRs in my-org/my-repo. Summarize changes, flag concerns, note approvals needed. | 0 9 * * 1-5 |
overnight-build | Use /steve to implement the next phase in /path/to/project | 0 2 * * * |
security-scan | Audit /path/to/repo for OWASP top 10 vulnerabilities. Report findings with severity. | 0 6 * * 1 |
dep-updater | Check for outdated dependencies in /path/to/project. Create a PR with safe updates. | 0 3 * * 0 |
FILE STRUCTURE
~/.claude/agents/
├── run-agent.sh # Executes claude -p, delivers results
├── manage-agent.sh # Create/list/remove/run-now/logs
├── schedules/ # Agent configs (JSON)
│ └── stock-news.json
└── logs/ # Execution logs
├── stock-news_20260211_080000.log
└── stock-news_launchd.log
~/Library/LaunchAgents/
└── com.claude.agent.stock-news.plist # macOS schedulerKEY DETAILS
- Agents run with
--dangerously-skip-permissions— fully autonomous, no prompts - Each agent gets its own launchd plist for reliable scheduling
- The Beeper delivery instruction is injected into the prompt, so the spawned Claude instance handles delivery itself via MCP
- Logs capture full output for every run with timestamps
- Agents can invoke other skills (
/steve,/rnv,/gsd:execute-plan) — it's a full Claude Code instance