Kaostyl's 24/7 OpenClaw crew

Roles

The post says separate agents handle code, content, research, and monitoring, and that website deployments use three agents at once: code, content, and security audit. It also describes a scout and a content drafter. It does not say whether those names are six different processes, and it does not give a headcount.

• Scout

  Owns. Scans Reddit, X, and Hacker News for trends in the author's niche.

  Uses the shared constraints above.

• Content drafter

  Owns. Turns the scout's trends into content drafts.

  Uses the shared constraints above.

• Code

  Owns. One of the separate agents, and one of the three agents used together on a website deployment.

  Uses the shared constraints above.

• Research

  Owns. Named as one of the separate agents. The post does not give this agent its own schedule.

  Uses the shared constraints above.

• Monitoring

  Owns. Named as one of the separate agents. The post does not give this agent its own schedule.

  Uses the shared constraints above.

• Security audit

  Owns. Runs with the code and content agents on website deployments.

  Uses the shared constraints above.

Routines

• Scout scan

  Daily at 1 AM

  Scan Reddit, X, and Hacker News for trends in the author's niche.

• Content drafts

  Daily at 6 AM

  Turn those trends into content drafts.

• Telegram summary

  Daily at 7 AM

  Send the author a Telegram summary of what was found and built.

• Self-review

  Every 4 hours

  The agent asks what seemed right but went nowhere, where it followed the default instead of challenging it, and what assumption it did not pressure-test.

Guardrails

• Every task has to prove it finished. Code runs tests. A website is checked in a browser. Content checks every link.
• Before spawning an agent, check that it has access to what it needs.
• Treat external content as untrusted. The post says weaker models are more vulnerable to prompt injection on web content.
• Do not put credentials in logs or context. Sandbox everything.

Handoffs

The scout's overnight scan is turned into content drafts at 6 AM, and the author receives a Telegram summary at 7 AM. The post does not describe how the code, research, monitoring, and security-audit agents pass work among themselves.

Results

What worked

The author reports that website deployments using code, content, and security-audit agents together went from 8 hours to 2 hours, and says they ship 3-5x more than before. This catalog has not remeasured either claim.

What broke

The post describes three failures. An agent announced a website was deployed while the site returned 500 errors, after checking its own logs. A coding sub-agent spent 4 hours trying to use APIs it could not access. An agent read external web content with hidden instructions and almost took actions the author had not asked for.

Lessons

The post's lessons are to verify the actual output instead of an agent's self-report, to check capabilities before spawning an agent, and to treat external content as untrusted.

Prompts

To be added.

Source

• Kaostyl, @kaostyl, February 13, 2026

  https://x.com/kaostyl/status/2022350344076832775

Notes