A growing catalog of production-grade long-running AI agents, built and running on Conductor.
Real autonomous agents are long-running, parallel, stateful, and failure-prone — they fan out across many sub-tasks, loop until done, call LLMs and tools, and must survive restarts without losing work. That is exactly what a durable workflow engine is for.
This repo is a community catalog of reference agent harnesses — each one a self-contained, runnable project that shows how to build a serious, production-grade agent on Conductor primitives. Clone one, read it, run it in minutes.
⭐ If Conductor powers your agents, star the repo → github.com/conductor-oss/conductor
The repository-level README is an index, not an operating guide for any one agent. Each ready harness owns its setup, safety requirements, workflows, and examples in its directory.
| Harness | Category | Status | Start here |
|---|---|---|---|
| security-harness | Security testing | ✅ Ready | README · Agent skill |
| coding-harness | Software delivery | ✅ Ready | README · Agent skill |
| deep-research | Research | 🚧 Coming soon | — |
| customer-support | Customer operations | 🚧 Coming soon | — |
Each harness is self-contained and may have different prerequisites, safety gates, and entry points. Choose one from the catalog and follow that harness's README; do not assume commands from one harness apply to another.
git clone https://github.com/conductor-oss/conductor-agents
cd conductor-agents/<harness>
cat README.mdIf an AI assistant is operating the harness, point it at that directory's SKILL.md.
Conductor was built at Netflix to run mission-critical workflows at scale — it's been open source ever since. When you build an agent on Conductor you get durability, parallelism, and observability that no custom orchestration layer can match:
| Need | Conductor primitive |
|---|---|
| Survive restarts mid-run | Durable execution — state lives in the server, not your process |
| Fan out across 100 sub-tasks | FORK_JOIN_DYNAMIC — parallel branches, automatic join |
| ReAct / tool-calling loop | DO_WHILE + LLM_CHAT_COMPLETE — native LLM tasks |
| See every decision & retry | Full execution history, replayable in the Conductor UI |
| Compose agents into pipelines | Sub-workflows, schedules, human-in-the-loop signals |
Learn more → conductor-oss.org · Star on GitHub ⭐
Some agents in this catalog use Agentspan — an open-source durable runtime for AI agents that compiles agent definitions directly into Conductor workflows. If you prefer a higher-level agent SDK over raw workflow JSON, Agentspan is the place to start.
Got a production-grade agent running on Conductor? We want it here.
A new harness needs:
- A top-level directory with a
README.md(hero line + a "run in ~30s" quickstart block) - A concise
SKILL.mdwithnameanddescriptionfrontmatter plus safe operating instructions - A
run.shentrypoint that auto-boots the stack - Conductor workflow JSON definitions
- Worker source in any language
Then:
- Add a row to the Harness Catalog table above.
- Add a matrix entry in
.github/workflows/ci.ymlso it runs under the shared quality bar (lint + tests). - Open a PR.
All production-grade, runnable agents are welcome. If it runs on Conductor and solves a real problem, it belongs here.