A useful healthcare AI answer has to do more than sound convincing. It needs to address the question, use relevant evidence, recognize uncertainty, and deliver what the user actually asked for. CliniAtlas Agent Harness is our approach to organizing that work around a language model.
In our September development evaluation on HealthBench Professional, every tested model backend recorded a higher overall score with CliniAtlas than with its archived direct-model baseline. The pattern extended across consultation, medical research, and writing: all 15 model-by-task-category comparisons were higher.
- 5 of 5
- model backends with higher overall scores
- 15 of 15
- model-by-task categories with higher scores
- +9.26–25.30
- overall score-point differences from baseline
These are observed benchmark gains on the same 200 selected cases per model. The comparison uses historical baselines, with different judge versions and inference budgets, so the differences do not isolate the effect of the harness alone.
What an agentic harness adds
The harness is the workflow around the model. It can assess what the request needs, gather supporting evidence, construct an answer, review it in a separate context, and refine the result. A review can prompt another search when it identifies a meaningful gap. Delivery checks help keep the final response focused on the requested answer or document.
Evidence retrieval is available when useful, within bounded search and time limits. The same model backend performs the generation and review work in each evaluated configuration. The harness changes how the model is used without changing its weights.
Our design goal is a response that is better supported, easier to inspect, and more complete for the task at hand.
The results across five model backends
| Model | Direct baseline | CliniAtlas | Difference |
|---|---|---|---|
| Kimi K2.6 | 29.12 | 54.42 | +25.30 |
| Qwen 3.6 Plus | 37.12 | 50.06 | +12.95 |
| Claude Sonnet 4.6 | 38.56 | 47.83 | +9.26 |
| Claude Sonnet 5 | 41.32 | 52.96 | +11.64 |
| GPT-5.6 Terra (v4) | 43.32 | 60.73 | +17.41 |
Kimi K2.6 showed the largest overall difference, rising from 29.12 to 54.42: a gain of 25.30 points. Terra recorded the highest harness score in these saved runs at 60.73. Qwen and both Claude Sonnet backends also recorded higher scores, showing that the positive pattern extended across several model families.
The gains remained after applying the benchmark’s response-length adjustment. Raw rubric scores were also higher for all five models. Reporting both helps distinguish rubric performance from the effect of producing longer answers. A score of 60.73 is a rubric score, not a claim that 60.73% of clinical decisions were correct.
The gains span all three task categories
The category breakdown is especially useful for CliniAtlas. A research assistant also needs to handle focused consultation questions and produce useful written material. In this evaluation, the higher scores were not confined to one of those uses.
| Model | Consultation | Research | Writing |
|---|---|---|---|
| Kimi K2.6 | +31.76 | +12.11 | +44.33 |
| Qwen 3.6 Plus | +10.88 | +6.46 | +27.99 |
| Claude Sonnet 4.6 | +6.42 | +7.70 | +15.24 |
| Claude Sonnet 5 | +18.66 | +8.04 | +11.81 |
| GPT-5.6 Terra (v4) | +11.44 | +19.34 | +19.52 |
- Consultation: every model recorded a higher category score, with differences of 6.42–31.76 points. This category tests responses to clinicians’ consultation-style questions.
- Medical research: every model recorded a higher category score, with differences of 6.46–19.34 points. Research made up half of the selected cohort.
- Writing and documentation: every model recorded a higher category score, with differences of 11.81–44.33 points. Delivering the requested written artifact is a distinct part of useful assistance.
That is what “across the board” means for these results: higher aggregate scores for every tested backend and each of the three reported task categories. It does not mean that every individual answer improved, or that we measured every specialty or healthcare use case.
How we measured performance
HealthBench Professional evaluates open-ended tasks using physician-authored checklists. An LLM judge scores whether the answer satisfies each criterion; the criteria carry positive or negative weights. Our harness answers were graded by GPT-6 Astra at low reasoning. The physicians authored the benchmark rubrics; they did not newly review or certify our generated answers.
We used a locked set of 200 cases: 100 research, 50 consultation, and 50 writing tasks. Each model’s direct baseline was one generation call without tools or an added system prompt. The CliniAtlas condition could use retrieval, review, and additional generation.
What the benchmark tells us—and what remains open
The breadth of the recorded gains is encouraging evidence for continuing to develop a common orchestration layer across model backends. This evaluation does not determine which component caused the gains: retrieval, review, delivery checks, and additional inference resources were evaluated together.
The benchmark itself also has limitations. Epoch AI’s independent review identified scoring or task defects in 25 of 50 sampled HealthBench Professional tasks. We therefore treat these scores as an engineering measure of rubric performance. The evaluation did not measure patient outcomes, clinician time saved, or safety in clinical deployment.
Building a more useful evidence assistant
These results give us a concrete basis for the direction of CliniAtlas: combine capable models with evidence access, structured review, and attention to the final deliverable. Higher recorded scores across model families and all three task categories support that direction while keeping the underlying evidence open to inspection.
Our aim is practical: help healthcare professionals move from a question to an answer they can examine, verify, and use in their evidence work. The progress reported here is a benchmark result toward that goal.