Run the agent harness as it ships

On a Claude subscription, the agent harness you run is Claude Code, and we think that's the right one to be on. A team gets more from setting down how work reaches the agent and how the result gets checked.

DeliveryCobble team2 min read
A line engraving of a curved concrete structure against a pale sky.
Photo by Jonny James on Unsplash

If your team pays for Claude through subscriptions, the agent harness you run is Claude Code. We think that's the right harness to be on anyway. A team's own effort pays off more in how work reaches the agent and how the result gets checked.

A harness is the code that turns a language model into an agent. The model can only write text, so the harness gives it tools, like reading a file or running a command, and runs the calls the model writes, passing back what came out.1

The harness runs the model's tool calls

One turn of the loop

The call

  1. ModelWrites read_file("invoices.ts") as texta tool call
  2. HarnessSpots the call and runs itruns the code
  3. Your machineOpens the file

The result

  1. Your machineThe file's contentsthe output
  2. HarnessPasses them backas text
  3. ModelReads them, writes the next call
The loop repeats until the task is done. Tool names differ between harnesses.

Anthropic's terms don't let other apps route requests through a Claude subscription. OpenCode, an open-source harness, took its Claude sign-in out on 19 March 2026.2 Another harness still works with an API key, but then you pay for every token on top of the subscription.

Claude Code itself had more than twenty releases in October 2026.3 A harness a team writes itself, or a third-party one like OpenCode, has to keep up with that pace. Anthropic also tunes Claude Code for its own models as they come out, so the official harness is usually the first to get the most out of a new one.

On one of our projects, each engineer set up their own way of handing work to the agent and checking what came back, with their own tools and prompts. The results varied with the engineer.

A team's process sits on top of the harness

Your team owns

  1. Which workEligible tickets from your trackerpicked up
  2. The briefWhat done means, from the ticket's authorbuilt to it
  3. The checkA second agent, then your merge rules

The maker provides

  1. HarnessClaude Code or Codex, unmodifiedruns
  2. ModelClaude or OpenAI's models, on your subscriptions
Cobble is how a team writes down the top lane. The harness and the model in the bottom lane come from Anthropic or OpenAI.

That's the part Cobble writes down once for the whole team. It runs Claude Code and Codex unmodified, on the subscriptions a team already has, because a harness of our own would mean chasing those releases ourselves and teams paying per token for it. Cobble takes eligible tickets from your tracker, and before the agent starts on one, it asks the ticket's author about whatever is missing. A second agent checks the work before the pull request opens.

Illustrative drawing of Cobble's run screen for ticket BIL-311, showing which agent ran each step: Shape in Claude Code at medium effort, Build in Claude Code at high, Check in Codex at high, Deliver in Claude Code at low, each done. A line below says the work was checked by a second agent, not the one that built it.
Photo: Cobble

If Anthropic ever let subscriptions run in any harness, or stopped shipping its own, we'd look at this choice again.

  1. Theo Browne's article "What is a 'harness'?" on X (23 October 2026) is the clearest longer explanation we know of, including how an app and its harness come apart. Link Back to text
  2. Anthropic, Claude Code documentation, "Legal and compliance", read 27 October 2026: it "does not permit third-party developers to offer Claude.ai login into their own applications, or to route requests through Free, Pro, or Max plan credentials on behalf of their users", and signing in to the unmodified Claude Code with your own subscription is allowed. OpenCode's pull request 18186, "anthropic legal requests", was merged on 19 March 2026. Anthropic, OpenCode Back to text
  3. Claude Code's releases on GitHub (anthropics/claude-code): 23 stable releases between 1 and 26 October 2026. Link Back to text

Read next

  1. If merging is your AI bottleneck, you're doing well

    Anthropic's merges grew from about 2,500 a day to over 30,000, and git couldn't keep up. We think where AI work piles up shows how far a team has got.

    Delivery3 min read
  2. More reviewers won't fix your AI code review bottleneck

    AI coding gains are real at the keyboard and mostly gone by release, and inside companies they stall at code review. Reviewers now send back nearly twice as many pull requests of the same size. More reviewers won't fix that. What arrives at review has to change.

    Delivery5 min read
  3. Don't let a model write your AGENTS.md

    An AGENTS.md written by your agent's setup command mostly repeats your docs and makes runs more expensive. The file is for the rules your senior engineers carry in their heads, each tied to the code it covers so it doesn't go stale.

    Delivery6 min read