Anthropic's $200 agent run bought 18 times the runtime of the $9 attempt

Divide each run by its clock. Nine dollars over 20 minutes is about $27 an hour ^[16]. Two hundred dollars over six hours is about $33 ^[17]. The working configuration cost roughly 22 times more and ran roughly 18 times longer ^[18], so the money mostly bought wall clock. The post attributes the working run to three coordinated agents ^[4], and those three agents ran at an hourly spend about a fifth above the single unscaffolded run ^[21].

Before that 22x goes into anyone's plan, three conditions have to hold. The task has to split into the roles a three-agent setup assumes ^[4]. The checks have to be machine-runnable, because the post ranks feedback, how the agent finds out whether what it did is right, as the part of the harness that contributes the most performance and receives the least attention ^[7]. And the 20-minute run has to be a genuine attempt at the same task. A run that stops after 20 minutes is cheap in part because it stopped. The post does not say what the task was or how many trials produced each figure ^[20].

The four-step progression the post cites from another team is more useful for planning, even though it ran on a different model. The instructions file was the biggest single step, 40 points off a one-in-five baseline ^[19]. Explicit verification commands added 20 more ^[19]. A progress template carried between sessions did not lift the ceiling; it held the result in an 80 to 100 percent band, with the same model in all four measurements ^[12].

Why a text file moves a number 40 points is mundane. A fresh session reads the repository. A convention that lives in someone's head, or in a Slack thread from three months ago, is not in the repository. "No es que la ignore, no puede verla," the post's author wrote of the agent: it is not that it ignores the rule, it cannot see it ^[14]. The recommendation he lands on is that everything that matters has to be written down, inside the repo ^[15].

The same logic is why I would read the leaderboard figure as a claim about someone else's repository. The post puts the best coding agents at 50 to 60 percent on SWE-bench Verified at the end of 2025, on curated tasks that already have tests, and calls that the most favourable possible setting ^[13]. Those tasks arrive with the verification a team would otherwise have to write itself.

A harness, in the post's definition, is everything in the engineering infrastructure around an agent that is not the model weights, so the complete agent is model plus harness ^[5]. The weights are trained by Anthropic, OpenAI or Google, and the team using them cannot open the model and adjust a parameter; the harness belongs entirely to that team and is ordinary software engineering, designed, versioned, tested and improved with data ^[6]. None of the five parts substitutes for the other four, and a perfect instructions file does little if the agent has no way to verify it followed them ^[8].