no-polling-agents
npx machina-cli add skill parcadei/Continuous-Claude-v3/no-polling-agents --openclawNo Polling for Background Agents
When launching parallel background agents, do NOT poll with sleep loops.
Pattern
Background agents write to status files when complete. Wait for them naturally.
DO
- Launch agents with
run_in_background: true - Continue with other work while agents run
- Check status file only when user asks or when you need results to proceed
- Trust the agent completion system
DON'T
- Run
sleep 10 && cat status.txtin loops - Continuously poll for completion
- Waste tokens checking status repeatedly
- Block on agents unless absolutely necessary
When to Check Status
- User explicitly asks "are they done?"
- You need agent output to proceed with next task
- Significant time has passed and user is waiting
Example
// Launch agents
Task({ ..., run_in_background: true })
Task({ ..., run_in_background: true })
// Continue with other work or conversation
// Agents will write to status file when done
// Only check when needed
cat .claude/cache/status.txt
Source
User feedback: "You can just wait until everyone pings you"
Source
git clone https://github.com/parcadei/Continuous-Claude-v3/blob/main/.claude/skills/no-polling-agents/SKILL.mdView on GitHub Overview
No Polling for Background Agents avoids sleep loops by letting background tasks signal completion via a status file. Launch agents in the background and continue work, checking status only when needed. This approach saves tokens and aligns with a natural completion system.
How This Skill Works
Background agents run with run_in_background: true and write their completion state to a status file. You proceed with other tasks and only read the status file when the user asks or when results are required to proceed. The status file (for example .claude/cache/status.txt) serves as the single source of truth for completion.
When to Use It
- You launch multiple background agents with run_in_background: true and want to continue conversing or working without blocking.
- The user explicitly asks, “are they done?” and you need a definitive answer.
- You need the agents’ output to proceed with the next task and cannot block on polling.
- Significant time has passed and the user is waiting, making polling wasteful.
- You want to minimize token waste from repeated status checks and rely on a single completion signal.
Quick Start
- Step 1: Launch agents with Task({ ..., run_in_background: true })
- Step 2: Continue with other work or conversation; do not poll
- Step 3: When needed, check status via cat .claude/cache/status.txt
Best Practices
- Launch agents with run_in_background: true to avoid blocking the main flow.
- Do not poll with sleep loops; rely on the status file written by agents.
- Check the status file only when the user asks or when results are required to proceed.
- Trust the agent completion system and its status-file signaling.
- Continue with other work while agents run to maximize throughput.
Example Use Cases
- Launch two tasks in the background using Task({ ..., run_in_background: true }) and continue the conversation while they run.
- When the user asks for results, read the status file at .claude/cache/status.txt to determine completion before presenting results.
- Avoid sleep-based polling; rely on the status file to signal completion and fetch outputs only when needed.
- If a long-running task is underway, keep the dialogue flowing and check status only when results are required.
- If time has passed and the user is waiting, use the status file to decide whether to present interim progress or final results.
