Surge: Opus at xhigh gained 11.3 points, while add-ons like Superpowers delivered at most 3

In recent months, feeds have been packed with add-ons for coding agents: skills, plugins, and orchestras comprising dozens of subagents. It sounds impressive, but does it actually help? The Surge team put it to the test with a benchmark and reached a conclusion many people will dislike: the most effective upgrade was also the most boring one.

How the test was conducted

They ran Opus 5.5 in Claude Code on 100 real-world engineering tasks from the sudo L7 benchmark. It evaluates more than whether “the tests passed”; it assesses staff-level engineering judgment: whether the agent found the root cause, checked adjacent scenarios, and avoided breaking nearby functionality.

The comparison included vanilla Claude Code, Superpowers (311K+ GitHub stars), Compound Engineering (25K+ stars), and the built-in /goal and /code-review plugins. Plain Opus at medium served as the baseline.

The results

• Superpowers: +2.4 points at about 1.1× the baseline cost.
• /goal + /code-review: +3.3 points at 1.8×.
• Compound Engineering: roughly 8× the cost and a statistically insignificant improvement.
• Opus at xhigh with no add-ons: +11.3 points at 3.5×.

The largest xhigh gains came in verification and thoroughness (+14.9) and persistence (+9.8). In other words, the model did not become “smarter”; it simply stopped settling for the first plausible fix.

An example from the report

A video editor crashes when the first clip is missing. At medium, the agent fixes the obvious location, declares success, and scores 14 out of 100 for $0.12. At xhigh, it notices that the fix redirects execution to a fallback where the clip is also assumed to exist, finds the second bug, and scores 100 out of 100 for $0.41. The price difference is tiny; the quality difference is enormous.

An even clearer example involved the task of “adding a Zoom link and passcode to the confirmation email.” One of the solutions—the authors do not specify which—loaded 37 additional skills and launched 17 agents: five to review the plan, five to simplify it, and seven to review the code. Not one noticed the central issue: the email is sent before payment, meaning people who have not paid would also receive the passcode. More agents do not equal more attention when they all examine the same task in the same way.

What to do in practice

1. Before installing another framework, increase the effort level. It is one setting, not a new dependency.
2. Calculate cost per task solved, not per run. Expensive xhigh is often cheaper than several rounds of revisions.
3. If you use skills, measure their contribution on your own tasks. GitHub stars are not a quality metric.
4. For critical areas such as payments and access control, do not rely on agent reviews. A human should verify the business logic.

Limitations

This is one benchmark, one model, and one execution environment. The tasks were selected by the vendor, and differences between configurations on small samples may be noise. Orchestration may perform better with other models or on long, multistep projects. In addition, xhigh costs 3.5 times more and may be excessive for routine edits.

But the direction is clear: start with the simplest lever. If the model makes mistakes because it is rushing, extra agents will not solve the problem; they will merely give you a larger bill and a feeling of control.