Swapping the scaffold moved Claude Opus 4.5 from 42% to 78% on CORE-Bench

Aider's `--auto-test` is the smallest working version of the argument. Switch it on and the tool runs your suite after each edit, feeds a non-zero exit back to the model, and tries again ^[7]. The sensor is the exit code, nothing more. Boeckeler, writing on Fowler's site, splits the wrap around a model into guides that steer before it acts and sensors that observe after so it can self-correct, and calls tests and linters the cheap sensors ^[8]. LangChain's formulation is shorter: "agent = model + harness" ^[5]. Anthropic describes the same loop as gather context, take action, verify results, and the Claude Code docs tell you to ask the model to write the tests, run them, and fix the failures ^[6].

Both benchmark deltas in the post are claims about someone else's repository. Between the two CORE-Bench scaffolds the spread is 36 percentage points, or 1.86 times the score, with the model held constant ^[11]. Before a number like that means anything in your tree, the command that decides whether a change worked has to exit non-zero when the change is wrong. It also has to run where the agent is, and the post's obstacles all sit in that second condition: this machine, this compose file, this `.env` that is not in git ^[14].

The LangChain figure is reported as a rank movement, and the post does not include the pass rates behind it ^[17].

"The first response is still wrong half the time," the post says of harness-driven work ^[13]. The harness changes who pays for attempts two through five: the model, against compiler output, instead of a human, and the human is usually the one who did not write the patch ^[13]. Once the output is meant to land in a shared repository, the post argues, "the chat is generating text about a system it cannot see" ^[12].

This is one practitioner's account, and the two scaffold results are third-party numbers it summarises ^[1][10]. The loop diagram survives that reading: the web-chat sequence runs prompt, text, paste, git, someone else, tests, so verification lands after the handoff ^[18]. In my view that ordering, not the model, is what makes chat-authored patches look finished and fail on boot. The post still recommends web chat for architecture, rubber-ducking, and first drafts of docs ^[15].