Get AI income methods before they spread.
Free weekly intelligence newsletter.
Research on AI-assisted software development shows single LLM coding achieves a 6.7% PR acceptance rate. Add a structured harness — a system that orchestrates multiple AI sessions with defined roles — and that number jumps to 70%. Archon v2 is the open-source tool that builds those harnesses, and the difference in output quality is not marginal.
What Is the Gap Between Vibe Coding and a Harness?
Vibe coding means giving an LLM a task and taking whatever it produces. It works for small, well-defined tasks — generating a function, writing a test, fixing a specific bug. At the PR level, where changes need to be correct, safe, and consistent with a codebase's architecture, a single unguided LLM produces code that gets rejected 93% of the time.
A harness adds structure. Different steps use different models. A planning step thinks through the approach before any code is written. A review step evaluates the output without access to the implementation reasoning. Each step has a defined role and a defined output format.
Stripe ships 1,300 AI-only PRs per week using this approach. The harness is what makes that scale possible.
How Does Archon v2 Build Harnesses?
Archon v2 lets you define workflows in YAML. Each node in the workflow is either a bash step or an LLM prompt. Nodes connect sequentially or with branching logic. The key capabilities:
- Per-node model selection — use a cheap, fast model for classification steps and a more capable model only where reasoning quality matters
- YAML workflow definition — the harness logic is a file you can version, share, and iterate on
- Bash + LLM mixing — deterministic steps (fetching data, applying labels, running tests) stay as bash; reasoning steps become LLM nodes
- Orchestration above Claude Code — Archon coordinates Claude Code sessions rather than replacing them
What Does a Simple Archon Workflow Look Like?
A basic issue-fixing workflow has four nodes:
- Bash: fetch the GitHub issue description
- LLM (fast model): classify the issue type and estimate complexity
- LLM (capable model): write the fix, following the governance rules loaded from CLAUDE.md
- Bash: open a pull request with the changes
That workflow, run on a real GitHub issue, produces a PR that a human reviewer can evaluate. The reviewer did not write the code — but the harness structured the process well enough that the output is reviewable.
Why Does Per-Node Model Selection Matter for Cost?
Not every step in a workflow needs the same model. Classifying an issue as a bug or a feature request requires less capability than writing a correct fix. Using a $0.14/M token model for classification and reserving a more capable model for implementation cuts the cost of running the workflow significantly without reducing output quality where it counts.
Per-node model selection is how harness-based development stays affordable at scale.
Archon supports OpenRouter, so any model available through OpenRouter is accessible per node. You define which model handles which step in the YAML config.
How Do You Get Started With Archon?
Clone the Archon repository from GitHub. Install dependencies with Bun. Configure your OpenRouter API key. Write a simple workflow YAML — start with two nodes: one bash step to read a file, one LLM step to summarise it. Run it. Once that works, extend to a real use case: issue triage, code review, or documentation generation.
Frequently Asked Questions
Where does the 6.7% versus 70% data come from?
The figures come from research on AI-assisted software development published in 2024 and 2025. The 6.7% figure reflects PR acceptance rates for single-agent, unguided LLM coding. The 70% figure reflects teams using structured multi-step harnesses with review steps. Cole Medin cites this data in his Archon v2 launch video.
Is Archon a replacement for Claude Code?
No. Archon sits above Claude Code and orchestrates it. Claude Code handles individual coding tasks within a workflow step. Archon handles the sequence of steps, the model selection per step, and the data passing between steps. You need both.
Can Archon workflows be shared across projects?
Yes. YAML workflow files are portable. A triage workflow built for one project works on any project with a similar issue structure. Teams can maintain a library of reusable workflow templates in a shared repository.
What is the minimum viable Archon setup for a solo developer?
One workflow covering the task you do most repetitively: issue triage, PR description writing, or test generation. Build that workflow, run it for a week, and evaluate whether the output saves you meaningful time. Add workflows only when the existing ones are stable and delivering consistent results.
Watch Cole demo Archon v2 live: Archon v2 Launch on YouTube
Get AI income methods before they spread.
Free weekly intelligence newsletter.
How to Get Your First AI Client Without Charging Anything