Paid computer-use agent arm across three reference verticals

July 21, 2026 ยท View on GitHub

Status: small-N local engineering evidence, not publication evidence.

On 2026-07-21, a paid claude-sonnet-5 computer-use agent was evaluated on three pinned, synthetic reference environments: Frappe Lending, openIMIS, and the pinned-local OpenEMR subset. Each trial began from a reset baseline, drove the same task contract used by the compiled reference, and was classified by an arm-independent system-of-record oracle. Pixels and the agent's self-report did not establish success.

Aggregate results

VerticalEnvironmentTrialsPrimary outcomeSilent incorrect successMean list-price model costMean wall time
InsuranceopenIMIS33/3 correct0/3$0.4793/run69.8 s
LendingFrappe Lending66/6 correct writes; 5/6 clean (one post-write cost-cap over-halt)0/6$0.4240/run67.3 s
HealthcareOpenEMR (local)60/6 correct; 6/6 missing write0/6$0.8901/run113.3 s

Across these 15 trials, the agent recorded zero silent incorrect successes. The only classified over-halt was the one post-write lending cost-cap halt; the insurance and OpenEMR arms recorded 0/3 and 0/6 over-halts, respectively. The OpenEMR result is a negative result: the bounded agent exhausted its action budget without completing the write in every trial.

Method and public/private boundary

The public, generic agent-baseline mechanism remains in openadapt_flow/benchmark/agent_baseline.py. It provides explicit paid-run opt-in, action and cost caps, usage accounting, and an agent self-report that is never trusted for scoring. The public benchmark schemas and aggregate failure taxonomy also remain available for independently authored tasks.

The application-specific driver adapters and oracle wiring, raw per-run JSON/JSONL, environment fingerprints, screenshots, and detailed cost ledger are private evidence. They are retained with file-level provenance and hashes in the private OpenAdaptAI/openadapt-corpus repository and are intentionally excluded from this public repository and its distributions.

Caveats

  • These agent trials used freshly provisioned baselines separate from the earlier compiled/API model-free subsets. The columns must not be presented as a matched three-arm comparison.
  • The sample is small: 3 trials for insurance and 6 each for lending and healthcare.
  • The environments used synthetic data on one local host. This is not customer, regulated-production, clean-machine, or broad reliability evidence.
  • A publication comparison still requires a preregistered, matched protocol with at least 10 fresh trials per task and condition, a common baseline, independent review, and complete reporting of incorrect success and over-halt.

This aggregate report does not claim publication readiness, certification, commercial availability, or superiority over computer-use agents generally.