AutoHarnessBench

Can frontier models automate agent harness R&D?

Each model is given a minimal harness scaffold, a task suite, and a 30-minute budget. It iteratively constructs and refines the harness by executing tasks and inspecting the resulting traces. The final harness is frozen and evaluated on held-out tasks.

Track
Benchmark
Actor

Benchmark suites

Harnesses are tested across four environments spanning function calling, expert-judged reasoning, tool use, retrieval, and simulated user interaction.

Function calling

BFCL Multi-Turn

Handle evolving requests by selecting and sequencing functions with valid arguments across multi-turn conversations.

Function calling 40 trusted tasks per actor

Healthcare reasoning

HealthBench Lite

Produce safe, useful healthcare responses graded against expert-written rubrics.

Rubric judged 100 trusted tasks Cross-model + self-improvement

Airline customer support

Tau three Airline

Resolve airline support requests using tools, policy, and a simulated customer.

Tool use 3 × 20-task trusted trials

Banking customer support

Tau three Banking

Handle banking requests using tools and retrieval over a 698-document knowledge base.

Retrieval + tool use 3 × 39-task trusted trials

Cite AutoHarnessBench

@misc{sleiman2026autoharnessbench,
  author       = {Essam Sleiman and Mersad Abbasi and Ayush Chakravarthy and Karina Nguyen},
  title        = {AutoHarnessBench: AI R\&D at the Harness Layer},
  year         = {2026},
  month        = jul,
  url          = {https://github.com/essamsleiman/autoharnessbench},
  note         = {GitHub repository}
}