Agentic and workflow benchmarks

5 tracked · snapshot reviewed September 28, 2026

These evaluations ask models to complete multi-step tasks in electronic health records, terminals, or administrative tools. Results depend on the model, its agent harness, available tools, and task setup. Configuration notes distinguish the systems and workflows being evaluated.

HealthAgentBench

Microsoft Research · 54 agentic tasks

Agent harnesses completing realistic terminal-based healthcare tasks built from real clinical artifacts.

CHI-Bench

actAVA · 75 operations workflows

Long-horizon healthcare operations workflows for agents: prior authorization, utilization management, and care management.

54.7%
eriusclaude-opus-5
37.3%
eriusclaude-opus-4-8
37.3%
claude-codeclaude-opus-5
33.3%
claude-codeclaude-opus-4-8
28.0%
claude-codeclaude-opus-4-6
26.2%
claude-codeclaude-sonnet-4-6
25.3%
codexgpt-5.6-sol
25.3%
openai-agentskimi-k3
24.4%
claude-codeclaude-opus-4-7
24.0%
claude-codeclaude-fable-5
22.7%
M
hermesMedGuard
20.9%
codexgpt-5.5

Showing top 12 of 44 indexed results. View all results.

PhysicianBench

academic team · 100 clinical tasks

Agents carrying out long-horizon physician workflows inside real EHR systems, verified by execution against those systems.

Separate evaluation setups. Compare results within each group; the graders and harnesses differ.

Anthropic evaluation · Opus 5 grader

Original paper · benchmark authors’ protocol

46.3 ± 1.2
GPT-5.5
31.7 ± 2.3
Claude Opus 4.6
29.3 ± 2.5
Claude Opus 4.7
27.7 ± 1.5
GPT-5.4
23.0 ± 2.6
Claude Sonnet 4.6
18.7 ± 2.9
D
DeepSeek V4-Pro
17.0 ± 2.6
Kimi-K2.6
16.7 ± 4.0
X
MiMo-v2.5-Pro
13.7 ± 4.0
A
Qwen3.6-Plus
8.7 ± 1.2
MiniMax M2.7
6.0 ± 1.0
Gemini Pro 3.1
5.3 ± 3.2
Grok-4.20

EHR-Complex

academic team · 3,915-task test set

Agentic clinical reasoning over MIMIC-IV records through SQL and Python, at patient and population level.

Showing top 12 of 18 indexed results. View all results.

HealthAdminBench

Kinetic Systems · 135 admin tasks

Computer-use agents completing healthcare administration workflows: prior authorizations, denial appeals, and DME ordering.

36.3%
Claude Opus 4.6 (computer-use agent)
26.7%
GPT-5.4 (computer-use agent)
15.6%
Kimi K2.5
14.8%
Claude Opus 4.6 (standardized harness)
13.3%
A
Qwen 3.5
11.9%
Gemini 3.1 Pro
5.9%
GPT-5.4 (standardized harness)

Which agentic and workflow benchmarks have results in this index?

5 as of September 28, 2026: HealthAgentBench (Claude Code (Opus 5): highest indexed score 55%); CHI-Bench (erius + claude-opus-5: highest indexed score 54.7%); PhysicianBench (Claude Opus 5.5 (max): highest indexed score 68.4%); EHR-Complex (GPT-5.4 (high reasoning): highest indexed score 0.65); HealthAdminBench (Claude Opus 4.6 (computer-use agent): highest indexed score 36.3%).

The other categories sit on the index: rubric-graded benchmarks, documentation and coding benchmarks, safety benchmarks, knowledge and exam benchmarks, composite indices.