Long-running autonomous coding agents - this is something I've always wanted to achieve. From the beginning, I put them inside an interval bash script loop. Now I have a better version using Claude Agent SDK in TypeScript called duyetbot-agent (need a better name!) that runs 24/7, written by Claude Code, but you can plan to build something similar with some ideas:
Build a long-running autonomous coding agent using Claude Agent SDK
that continuously processes tasks, creates PRs, and manages its own backlog.
Core components:
- Orchestrator: Main loop managing session lifecycle, scheduler, monitoring, analytics dashboard
- Sub-Agents: Planner Agent, Coder Agent, Reviewer Agent, PR Manager
- Tools: Git worktree management, GitHub MCP, memory persistence (Supermemory)
- Stop Hooks: Graceful exit with state preservation
- Skills: Reusable workflow templates
- Configuration: Flexible YAML-based settings
- Quality Gates: Automated checks before PR creation
...
I was thinking of publishing it for general use, but it still needs a lot of work to be generic enough. (2026-01-25: Yes someone did it, check out OpenClaw (Clawdbot → BoltBot → OpenClaw, they changed the name a few times) - a personal agent and I am really impressed, you might try it out). Don't expect too much from me though, some big name will probably do it first. I believe this is a milestone the industry has always wanted to reach before we get to AGI.
Claude Code + Ralph Loop is the easiest way to try this, it uses StopHook to extend your session and keep working for hours or days to complete your task. See Boris's explanation:
Claude consistently runs for minutes, hours, and days at a time (using Stop hooks) https://x.com/simonw/status/2004916070973645242
I've also put my agent source code folder under its loop, to build a self-improving coding agent. I want to see how possible it is to have an agent that codes itself. It's not as good as expected with many issues, but it's still something I want to achieve
- bad execution loop and bad plan will generate broken agent and stops infinite loops
- reflection mechanism, need to extract the failure, error pattern, root cause to enhance the codebase itself
- better context engineering and long-term memory improvement
- code safety, detect potentially dangerous operations, monitoring and rollback, etc
Some solutions out there you can try: Continuous Claude, Continuous-Claude-v3
Good reads:
- Scaling long-running autonomous coding
- Anthropic CEO on Claude, AGI & the Future of AI & Humanity | Lex Fridman Podcast #452
- 2025 LLM Year in Review - Andrej Karpathy
Next: Claws
Claude Code + Ralph Loop
The ralph-wiggum plugin is my favorite for long-running tasks or vibe coding on fun projects while I'm asleep. You define a goal condition and let the agent loop until it verifiably reaches that goal. With cheap Z.AI GLM 4.7 tokens, I can let it run 24/7. Run it with --permission-mode=dontAsk or --dangerously-skip-permissions.
/plugin install ralph-wiggum@claude-plugins-official
/ralph-wiggum:ralph-loop "Implement feature X following TDD:
1. Write failing tests
2. Implement feature
3. Run tests
4. If any fail, debug and fix
5. Refactor if needed
6. Repeat until all green
7. Output: <promise>COMPLETE</promise>" --completion-promise "COMPLETE"
Claude cron scheduling tools
Claude Code now has native scheduled tasks built in — this can replace the loop or Ralph Loop for many use cases, and I find it works much better. The /loop command lets you schedule recurring prompts that fire in the background while you keep working.
Tasks are session-scoped: they live in the current Claude Code process and are gone when you exit. For durable scheduling that survives restarts, use GitHub Actions (see below) or Desktop scheduled tasks.
/loop 5m check if the deployment finished and tell me what happened
/loop babysit all my PRs, fix issues based on code review comments, merge when they look good enough, every hour
You can also loop over other commands or skills:
/loop 20m /review-pr 1234
Ask Claude in natural language to list or cancel tasks, or reference the underlying tools directly:
what scheduled tasks do I have?
cancel the deploy check job
Under the hood, Claude uses CronCreate, CronList, and CronDelete tools. Each task has an 8-character ID, and a session can hold up to 50 scheduled tasks. Recurring tasks auto-expire after 3 days. Set CLAUDE_CODE_DISABLE_CRON=1 to disable the scheduler entirely.
Paperclip
If OpenClaw is the employee, Paperclip is the company. For zero-human companies, Paperclip lets you spin up an entire org where every role - CEO, staff engineers, reviewers - is a Claude Code instance, from the smartest model down to the cheapest. I'm using it now to run Duet Company: one Opus agent as the CEO, a squad of cheap GLM 4.7 workers burning through tickets, and whatever free model is available as the reviewer - Gemma 4, Qwen 3.6. I defined the org chart, set some goals and cronjobs, and it just starts going - assigning tasks, breaking them down, discovering more work on its own. It's wild.
It looks like a real enterprise now - the agents keep reassigning tickets back and forth to each other all day. I accidentally built corporate bureaucracy in AI form. At least they don't schedule unnecessary meetings... yet.
Watch the Duet Company demo video
Claude Code (+ OpenRouter) on GitHub Actions
Something else you can try for maximize automation that Claude Code Action which is Claude Agent SDK that running on Github Actions. The best part is I'm running Claude GitHub Actions with OpenRouter at no cost by using free models. I have an OpenRouter preset that can switch between SOTA free models automatically.
- name: Run Claude Code Review
id: review
uses: anthropics/claude-code-action@v1
env:
ANTHROPIC_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
ANTHROPIC_BASE_URL: https://openrouter.ai/api
ANTHROPIC_DEFAULT_HAIKU_MODEL: xiaomi/mimo-v2-flash:free
ANTHROPIC_DEFAULT_SONNET_MODEL: xiaomi/mimo-v2-flash:free
ANTHROPIC_DEFAULT_OPUS_MODEL: xiaomi/mimo-v2-flash:free
with:
anthropic_api_key: ${{ secrets.OPENROUTER_API_KEY }}
additional_permissions: |
actions: read
claude_args: |
--allowed-tools Bash Edit Glob Grep Read Write
--mcp-config .github/mcp-config.json
plugins: |
ralph-wiggum@claude-plugins-official
I put together some reusable workflows at duyet/github-actions that other repos can reuse:
name: Claude Code Review
on:
pull_request:
types: [opened, synchronize]
jobs:
review:
uses: duyet/github-actions/.github/workflows/claude-code-review.yml@main
permissions:
contents: read
pull-requests: write
issues: read
id-token: write
secrets:
api_key: ${{ secrets.OPENROUTER_API_KEY }}
bot_github_token: ${{ secrets.DUYETBOT_GITHUB_TOKEN }}
Check out the official documentation: Claude Code GitHub Actions. Some use cases:
- Code Review - Automated PR reviews with AI feedback
- Nightly Codebase Analysis - A scheduled workflow that scans the codebase every night, finds things to improve or refactor, creates an issue, and assigns it to @claude to fix via PR
- Triggering cross-repo workflows (e.g. SDK change -> updates docs).
This way you can have Claude Code + OpenRouter free or cheap models running 24/7 for you. A lot of automation becomes possible: smart cronjobs, automated refactoring, documentation sync, etc. The AI does the boring stuff while you sleep.
