Within each matched comparison

Held fixedCoding task
→
Held fixedUnderlying model
→
MeasuredSuccess & cost
The harness changes the route between request and result
Harness part
Tested choice
Observed role
Memory
Remove stale tool output first; summarize older history only if still needed.
Tight context: chiefly kept runs from ending when history overflowed.
Planning
Give the model a persistent task list it can update.
Model-dependent: weaker models reached edits more often; stronger ones did less redundant checking.
Action interface
Structured file tools or direct Bash commands.
Skill-dependent: structure helped weaker Bash users; capable Bash users worked at lower measured cost.

The action-interface comparison bundled tools with different instructions and safeguards, so it does not isolate tool count. Results describe four tested models and two benchmarks, not every coding assistant.