Benchmarks

April 30, 2026 · View on GitHub

HarnessX benchmark adapters — each sub-directory is a self-contained adapter for one evaluation suite.

DirectoryBenchmarkStatus
terminal_bench_2/Terminal Bench 2.0Available
gaia/GAIAAvailable
swebench/SWE-benchAvailable
tau2/TAU2-BenchAvailable
evoclaw/EvoClAWOngoing
locomo/LoCoMoOngoing
osworld/OSWorldOngoing

Benchmark descriptions

Terminal Bench 2.0

Terminal Bench 2.0 tests agents on 89 bash/file-system tasks executed inside Harbor-managed containers. Tasks cover compilation, debugging, system administration, cryptography, and ML workloads. Use it to evaluate low-level terminal proficiency and tool-use reliability.

GAIA

GAIA tests general agent capability across reasoning, tool use, and web information retrieval. Tasks range from simple factual lookups to multi-step questions requiring file parsing, web search, and cross-source synthesis. Use it to evaluate how well a harness handles open-ended, real-world assistant tasks.

SWE-bench

SWE-bench tests software engineering capability using real GitHub issues from popular Python repositories. Each task asks the agent to produce a patch that resolves the issue and passes the associated test suite. Use it to evaluate coding and debugging ability in a realistic, large-codebase setting.

TAU2-Bench

TAU2-Bench tests tool-augmented agents in user-simulation scenarios. Tasks involve multi-turn interactions with simulated users and structured tool APIs (retail, airline, etc.), evaluating whether the agent can complete user requests correctly while following domain-specific constraints.


Results

Terminal Bench 2.0

89 terminal/coding tasks inside Harbor-managed containers.

AgentModelTasksPassScoreRuns per taskNotes
HarnessXclaude-opus-4-6895663.0%1 (k=1)OpenSandbox
Claude Codeclaude-opus-4-68958.0% ± 2.95 (k=5)tbench.ai
HarnessXclaude-haiku-4-5892831.5%1 (k=1)OpenSandbox
Claude Code 2.0.31claude-haiku-4-58927.5% ± 2.85 (k=5)tbench.ai

HarnessX was run once per task (k=1); the official leaderboard uses k=5 and reports mean ± confidence interval. Direct comparison is not strictly valid, but the scores are closely aligned.

TAU2-Bench

Multi-turn customer-service simulation across three domains. Agent: Sonnet 4.5 + Extended Thinking (budget: 62976 tokens). User simulator: GPT-5.2. The w/o HarnessX column is the published tau2 leaderboard score for the same model.

DomainTasksPassPass ratew/o HarnessXNotes
telecom1149886.0%84.9%base split
airline503774.0%72.0%
retail1148171.1%72.4%StopGuard enabled

StopGuard strips premature ###STOP### tokens from GPT user-simulator confirmation messages, recovering tasks that would otherwise fail before the agent executes the write tool (~8 pp on retail without it).