What happened
Surge AI's November 5, 2026 announcement describes an experiment using Opus 5.5 in Claude Code on 100 tasks from its sudo L7 benchmark. The benchmark uses real codebases and expert-written rubrics that evaluate more than functional correctness, including verification, practical judgment, and communication. Surge held the model fixed while comparing workflow configurations, asking how much improvement came from changing the harness around it.
The configurations included vanilla Claude Code; Superpowers, a skills plugin that enforces planning and testing discipline; Compound Engineering, with planning, work, review workflows and specialized reviewer agents; and Claude Code's built-in /goal and /code-review plugins. Surge also ran vanilla Claude Code at xhigh thinking effort, compared with medium effort for the harness comparisons. Its headline result: xhigh effort improved the score by 11.3 points over medium, while the most effective harness additions improved scores by 2 to 3 points.
Superpowers gained 2.4 points at roughly 1.1 times the cost of vanilla Claude Code. The /goal and /code-review combination gained 3.3 points at 1.8 times the cost. Compound Engineering cost eight times as much as vanilla, without a statistically clear improvement in this experiment. Surge says the xhigh run improved rubric scores by 14.9 points on Verification & Thoroughness, 9.8 on Persistence, and 7.0 on Thought Partnership.
The post illustrates the distinction with a video-editing task. Asked to make the first clip optional, the medium-effort agent fixed the immediate crash but missed a fallback path that also assumed the clip existed; it scored 14/100 and cost $0.12. The xhigh agent traced the path, found the second issue, scored 100/100, and cost $0.41. In another example, a configuration with 37 added skills spawned 17 agents to review a change involving Zoom links in event-confirmation emails. None caught that the email went out before payment, which could expose the passcode to unpaid registrants.
Surge recommends vanilla Claude Code as a low-cost baseline, Superpowers for an inexpensive improvement, /goal followed by /code-review for more thorough work, and higher thinking effort when quality matters more than cost. It advises teams to benchmark elaborate harnesses on their own tasks. The post says Surge has not yet tested combining xhigh effort with the plugins or comparing configurations under identical dollar or time budgets.
What the archive adds
New Runtime's earlier coverage described a coding-agent harness as an execution loop surrounded by infrastructure (April 23, 2026 note). The present experiment adds comparative measurements of workflow changes around a fixed model; it is a new fact on a known topic, not a restatement.
Trend signal
This supports the idea that harness architecture can outlast model choice: Surge isolates the workflow around Opus 5.5 and measures its effects. The result also cautions that more elaborate harnesses are not automatically better; the tested gains depended on the specific configuration and tasks.
Why it matters
Teams often treat planning agents, reviewers, and skills as an obvious route to more reliable coding. Surge's results suggest a more disciplined sequence: establish a baseline, try a lower-complexity change such as higher effort, then measure additions against real tasks. That matters because an orchestration layer can add cost and maintenance while leaving the important failure mode untouched. The 17-agent example is not proof that multi-agent review is useless; it shows that reviewers can miss a central product requirement when their work is organized around the wrong assumptions.
The scores also argue against measuring coding agents only by whether tests pass. Verification, persistence, and judgment can distinguish a patch that fixes the visible symptom from one that follows the code path far enough to catch the next failure. For teams, the practical unit of comparison should therefore include task quality, cost, and the kinds of errors that matter in their own product—not agent count or workflow complexity.
Limits and uncertainty
These are Surge's results on one model, Opus 5.5, one benchmark distribution, and particular harness configurations; they do not establish that xhigh effort wins across models or workloads. The post reports no statistically clear Compound Engineering improvement, not proof that it never helps. The Zoom-passcode failure is one example, not a general measurement of multi-agent review. Surge says it has not tested the combination of xhigh with /goal and /code-review or compared the setups under equal dollar or time budgets. Independent replication is not established by the announcement.
Reader takeaway
Treat a more elaborate coding-agent harness as a hypothesis to test, not an upgrade by default. On Surge's benchmark, raising thinking effort yielded the largest measured score increase, while some add-ons delivered smaller gains at higher cost. Run the comparison on tasks your team actually cares about, including checks for verification and product judgment, before committing to extra orchestration.
Open question
Would xhigh effort combined with /goal and /code-review improve results enough to justify the additional cost under an equal-budget comparison?