# Agent9 Clinic Batch Execution - Owner: Managed Clinic Provider Executor package and Agent9-backed Evidence materialization - Source: TCID-13070 Agent9 replay `f2fc8501-bf47-42cd-9074-3d3653551718` - Implementation status: deployed to Agent9 internal production and the Agent9-backed Auto Oncall branch runtime - Effective status: `optimized_verified` ## Bounded problem The prior replay completed in about 19 minutes 30 seconds and made 77 Managed Clinic calls. Each call created a separate Tool Sandbox even though queries in one Evidence round were independent and the Clinic contract already authorized concurrent execution inside one deterministic batch. ## Intended change One Evidence round sends at most twelve query entries in one Managed Clinic call. The Tool executes at most six entries concurrently and preserves one `provider-result-v1` success or safe error per stable `queryId`. Corrections are separate same-Turn batches containing only changed failed entries. The runner retains final-outcome materialization, bounded attempt metadata, query coverage, canonical evidence references, the four-correction ceiling, and Scenario refinement semantics. ## Verification Managed Clinic v16 package digest `16e0dc05209751a82a7a17e89d68f2a2ae1707f6212dff2528a33e993f0d8df6` was loaded by Agent9 internal production and installed on the Auto Oncall Evidence Agent at installation revision 7. The branch runtime image `448187b4d60b0d612b8a09871ac716c5a3dcb1c0` was read back Ready with zero restarts before the read-only replay. TCID-13070 replay `fd3b2042-506f-4da8-9a60-e0b0bd293bf3` completed with HTTP 200 in 810 seconds and Judge verdict `ai_better`. Three initial Evidence rounds and two authorized refinement rounds executed as five Managed Clinic calls containing 56 independent query entries. All 56 entries materialized a bounded result, no correction batch was needed, and final `execution.evidenceAcquisition` was `complete` with no failure codes. The prior baseline replay `f2fc8501-bf47-42cd-9074-3d3653551718` took about 1,170 seconds and created 77 Managed Tool calls. Scenario selection is stochastic, so the 31% total-wall reduction is supporting evidence rather than an isolated batch benchmark. The direct transport criterion is the reduction from one Sandbox per query to one Sandbox per Evidence round while preserving all per-query coverage and errors. ## Next validation action Monitor later Agent9-backed Clinic replays for batch-level transport failures or query starvation. Treat a query-specific Provider rejection independently; it must not discard successful peers from the same batch.