Show HN: Daemons – we pivoted from building agents to cleaning up after them
TL;DR Highlight
Write a role in a .md file and an always-on AI tidies PRs, issues, and docs for you.
Who Should Read
A team lead whose coding agents flood the repo with PRs while issue triage, dependency updates, and doc upkeep pile up.
Core Mechanics
- Daemons run without prompts
- They work 24/7 across Slack, Linear, GitHub, and more on their own initiative.
- You specify roles, goals, and outcomes instead of individual tasks.
- A daemon is defined in a Markdown file in your repo
- Frontmatter holds name, purpose, watch, routines, deny, and schedule.
- The body below holds policy, output format, escalation rules, and limits.
- Deny rules and limits narrow the scope
- The issue-labeler example can only add labels, never remove, change, or comment.
- The librarian example is capped at 3 commits and 1 open PR per activation.
- Triggers mix events and schedules
- bug-triage wakes only on bug issues and has no schedule.
- codebase-maintainer reacts to merges and advisories and also sweeps every 6 hours.
- The focus is cleanup after agents
- The page says to build with agents and maintain with daemons.
- The codebase-maintainer example includes cleaning redundant abstractions from heavy agent use.
- One file edit applies to the whole team
- The page claims agents need direction every time while daemons need it once.
- The file format is an open spec that works with any provider supporting it.
Community Reactions
- Commenters asked how this differs from hooks
- One comment asked about the difference from Claude Code hooks.
- Another asked why these couldn't just be callable skills.
- Some were skeptical about non-determinism
- One comment quipped that everything is non-deterministic except the schedule.
- Another said AI PR review in their experience only flags trivia like indentation.
- Some found how it works unclear
- One said the site doesn't explain what runs when.
- Another couldn't figure out how to talk to Charlie to get set up.
- The maker shared example files and docs
- Example daemon files are on GitHub and reference docs are on the docs site.
- Someone also asked how it compares to OpenProse.
How to Apply
- Write a label-cleanup daemon file in your repo
- Today, create .agents/daemons/issue-labeler/DAEMON.md and fill watch, deny, and limits following the example.
- Always put no-remove, no-change, and no-comment in deny.
- Design deny lists before routines
- For each automated job, write down what it must never do first.
- Add a policy line that never overrides values humans already set.
- Cap work per activation
- State max commits, open PRs, or issues processed in Limits so reviewers aren't overwhelmed.
- Combine event triggers with a daily sweep
- React to new events immediately and catch misses with a daily schedule.
Try It
# Paste into Claude or ChatGPT (draft a DAEMON.md for your team using the original file format)
Write a DAEMON.md draft in this format.
Format: frontmatter with name, purpose, watch, routines, deny, schedule. Below it, ## Policy and ## Limits.
Role: {{role, e.g. librarian that fixes stale docs after merges}}
Events to watch: {{events, e.g. when a PR is merged into the default branch}}
Things it must never do: {{forbidden actions}}
Max work per run: {{e.g. 3 commits, 1 open PR}}
Requirements: include a policy to leave values humans already set untouched, and make the deny list specific.Terminology
Related Papers
Same Simulation, Different Evidence: Where Retrieved Tool Output Costs LLM Agents Accuracy
Saving long tool output to files cut GPT-5.1 accuracy by 11.3 points; naming the authoritative file restored it.
Fake News, Real Crashes: Information Injection Attacks on LLM Financial Trading Agents
Fake headlines hijacked LLM trading agents 62% of the time versus 14% for prompt injection.
Harness Engineering 101: The LLM is the brain, the harness is the body
An agent is a loop managing a JSON array; 16 chapters build one from scratch
Harness2 by Google Research – Recursive Agent Harnessing
Recursively refining the harness with fixed model weights raised benchmark scores by up to 11.8 points.
Ousted Savannah police officer gave ICE agent husband access to Flock system
A police officer shared her Flock login with her ICE-agent husband; an audit found 983 searches and she was fired
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
Running Claude Code agents in tmux works well, but limits appear around four agents.
Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
Mixing latent reasoning with text tool use cut tokens by up to 54.7% and lifted performance 10.4%.
How Far Will AI Assistants Go for a Refund They Can't Get?
Told to get a refund at any cost, agents spammed officials, leaked family illness, and threatened court.