Pimp My IDE / Garage Dispatch
← Back to garage
September 17, 2026 · agent harnesses / matched comparisons / operating cost

The model is not the whole machine.

A coding model never arrives naked. It arrives bolted to a prompt, tool protocol, retry policy, context strategy, permission system, terminal, and receipt trail. Benchmark the assembled car—or stop pretending the engine badge tells you how it drives.

THE TAKE: “same model” is not “same agent.” Treat the harness as a versioned production dependency. Compare it on matched work, with the environment and budget held still, then keep enough traces to explain the gap.
Put the comparison on the swap stand ↓

The leaderboard forgot the transmission.

HarnessTax is a useful correction to model-only shopping. Its dashboard pairs the same named model across Pi, Claude Code, and Codex harnesses on SWE-bench Lite and Terminal-Bench 2, then shows success and cost together. The important word is pairs: model, task cohort, and repetitions are held close enough to inspect the harness effect instead of comparing unrelated headline scores.[1]

One published SWE-bench pairing for Claude Fable 5 covers 30 tasks over three repetitions. Claude Code versus Pi showed no clear success-rate separation after the dashboard’s multiple-comparison treatment, while the reported mean cost ratio was about 2.0 with a 95% interval of 1.76–2.33. Codex versus Pi on the same pairing likewise showed no clear success-rate difference, with a reported cost ratio around 1.34.[2]

That is not a universal harness ranking. It is a better question: what did this assembly buy on this matched cohort, and what did it burn? The Terminal-Bench comparisons are explicitly labeled descriptive common-30 estimates without confirmatory tests. Keep that caveat attached; a dashboard is not permission to sand off uncertainty.

The harness can spend more, retry differently, expose better tools, or fail more gracefully—and every one of those choices can masquerade as “model quality.”

The IDE is becoming a harness socket.

VS Code 1.138 makes the assembly visible. Its Agent Host can run Codex with Copilot- or ChatGPT-backed models, continue a session across apps, attach a quick chat to a folder or isolated worktree, use built-in, extension, and MCP tools, and run local sessions in a Dev Container. That is not one feature. It is a matrix of model source, session host, tool surface, workspace attachment, and runtime placement.[3]

The upgrade is flexibility. The operating hazard is silent configuration drift. If a result changes after switching surface, subscription, workspace, or execution floor, the model name is the least interesting line in the incident report.

Release notes are dyno sheets.

Claude Code 2.1.274 is packed with harness-level behavior: a bounded first-turn MCP startup wait, self-healing for some corrupted transcripts that caused endless retries, longer per-server MCP timeouts, context compaction fixes, preserved agent flags after auto-update, and leaner code-review prompting for models without tuned settings. None of those changes alters the weights. All can alter latency, cost, completion, or failure shape.[4]

This is why “we tested Model X” is weak release evidence. Record the harness version, resolved configuration, connected tools, environment identity, turn and retry ceilings, and exact acceptance oracle. When the behavior moves, diff the chassis before blaming the cylinder.

Good harnesses manufacture disagreement.

Cloudflare’s open security-audit skill makes the control logic concrete. It separates reconnaissance, coverage-led hunting, candidate validation, structured findings, independent record verification, and reporting. The finder does not certify its own finding. Confirmed, needs-validation, and rejected are separate states. Validators run again after material replacements.[5]

That architecture matters more than the swagger of “multiple agents.” Parallel agreement can be correlated enthusiasm. A useful harness assigns different jobs, preserves rejected candidates, validates machine-readable records, and gives uncertainty somewhere honest to live.

Buy the lap, not the logo.

For an adoption decision, run the same tasks with the same model build, same environment, same permissions, same task order, and the same stopping rules. Price the whole rollout, not just successful calls. Capture success, wall time, tokens or spend, turns, tool errors, retries, changed files, and the final oracle. Repeat enough to see variance.

The swap stand below replaces a model dropdown with a comparison mode plus four parity locks. It executes no benchmark and invents no telemetry. Its readiness dial is a teaching proxy; closing every lock produces a review gate ready plan, not a winning harness.

Interactive makeover / matched-agent evaluation

Harness Swap Stand.

Traditional purpose replaced: pick an agent from a leaderboard. Better version: choose why you are comparing, then lock engine identity, work, environment, and budget before reading the result. Native radios and checkboxes preserve keyboard and screen-reader operation.

Choose the inspection lap

The carriage changes the receipt, not the evidence standard.

Comparison purpose

Position indicator: mirrors the selected native radio.

Matched-comparison locks

Comparison readiness

MATCHED COMPARE · 0/4 locks closed · PLAN INCOMPLETE. A model label alone cannot isolate the harness.

Open the five-source evidence drawer
[1] HarnessTax, generated September 16, 2026 — matched same-model comparisons across Pi, Claude Code, and Codex on SWE-bench Lite and Terminal-Bench 2. [2] HarnessTax SWE-bench harness-effect data — task counts, repetitions, success deltas, uncertainty, multiple-comparison labels, cost ratios, and per-task breakdowns. Direct machine-readable source. [3] Visual Studio Code 1.138 release notes, September 16, 2026 — Agent Host Codex expansion, cross-app continuation, VS Code/MCP tools, quick-chat workspace conversion, worktrees, and local Dev Containers. [4] Claude Code 2.1.274 release notes, September 17, 2026 — MCP startup/timeouts, transcript retry recovery, context compaction, update flag preservation, and review-prompt changes. [5] Cloudflare security-audit skill, repository read September 17, 2026 — six-phase audit harness, isolated hunting, independent verification, structured verdicts, and validators. Discussion trail: Hacker News item 49733726, “HarnessTax: How Much Does the Harness Matter for Coding Agents?”

Source boundary: the concrete cost ratios above describe one named, matched benchmark pairing. They do not establish a universal ranking. Terminal-Bench panels identify their common-30 comparisons as descriptive. VS Code and Claude Code sources are vendor release notes; Cloudflare documents its own audit workflow.