Specialised agent harnesses need a fallback route
In short: OO-Harness gives specialist agent code a regression floor: preserve the base result unless the current failure matches a tested repair.
↗Li et al. — Object-Oriented HarnessOn 20 October 2026, Google Cloud AI Research and UIUC researchers reported a 15.79-point AutomationBench gain from OO-Harness: 77.41 mean partial credit against 61.62 for the base. The same frozen Gemini 3.5 Flash executor sat underneath both.
The useful result is narrower than “specialise the harness”. Specialised agent harnesses need a fallback route. OO-Harness runs the base first, inspects that trajectory, and sends the task to a specialist only when the evidence matches a learned repair. Otherwise it keeps the base result.
OO-Harness preserves a working base result
The paper evolved subclasses on discovery tasks, froze the hierarchy, then evaluated it on held-out work. It reported the strongest primary score among its compared methods on three benchmarks.
| Held-out benchmark | Base | OO-Harness |
|---|---|---|
| AutomationBench mean | 61.62 | 77.41 |
| Workspace-Bench mean | 25.45 | 39.48 |
| Toolathlon-GYM strict pass | 31.15 | 37.45 |
Reported percentages are means across three executor runs; I did not reproduce them.
The routing case study is more useful than the leaderboard. Among 300 AutomationBench tasks, 134 reused the base result. The other 166 went to specialists after their base trajectories scored much lower on average. In a separate 49-task test with unfamiliar application families, the router chose the base every time.
My fallback-route receipt for a specialist harness
This is editorial judgment, not a Trigent or client deployment. I would let a specialist override the base only when the run records five things: the observed failure signature, specialist version, validation set, preserved base result and total extra-run cost.
That receipt makes the decision reviewable. “Spreadsheet task” is a label. “The saved workbook contains unevaluated formula strings, and specialist v3 fixed that failure on the held-out set” is routing evidence.
OO-Harness still leaves a cost and replication gap
The paper's step count excludes auxiliary selection calls and unselected rollouts, so it is not a complete cost measure. Its routing and inheritance ablations are single runs. I also did not verify a public implementation or an independent reproduction.
Those limits do not erase the design lesson. They set the acceptance bar: measure the whole route, and keep the general path available when specialisation has no matching evidence.
What's in it for you
- Keep the base output until a tested specialist earns the override.
- Log why the router switched paths and what the switch cost end to end.
Specialise by observed failure, and preserve the base as the regression floor.
