The Best Enterprise Agent Scores 0.663 , on Tasks Built from Real Work

Every major agent benchmark in production use today was designed on a desk, not extracted from a desk. GAIA, WebArena, AgentBench: all synthetic task construction, all reasonable approximations of real work, all missing the specific failure modes that appear when agents touch actual enterprise files, actual tool chains, and actual business deliverables. The assumption baked into those benchmarks is that synthetic tasks are a good proxy for the real thing. EnterpriseClawBench tests that assumption directly, and the answer is no.

The benchmark starts with a proprietary archive of real workplace agent sessions. Each session gets processed into a reproducible task: fixtures are recovered, prompts are rewritten to strip internal identifiers, role classes and skill subclasses are assigned, and evaluation is paired with both hard rules and semantic rubrics. The result is 852 tasks that are grounded in what enterprise agents actually do, not what researchers assume they do. That distinction matters structurally. Synthetic benchmarks tend to over-represent clean, well-scoped tasks. Real workplace sessions contain heterogeneous file types, ambiguous instructions, multi-step artifact delivery requirements, and implicit quality standards that no prompt engineer would think to encode from scratch.

The construction protocol also changes what gets measured. Instead of collapsing performance into a single accuracy score, EnterpriseClawBench tracks harness-model combinations, artifact delivery completeness, visual quality of outputs, cost per task, runtime, and skill-transfer behavior across role classes. That reporting surface is wider because the tasks demand it. A session that produces the right spreadsheet but at three times the expected cost and with broken chart formatting is not a pass. For teams evaluating agents before production deployment, the takeaway is direct: if your current eval stack does not report artifact quality and cost alongside task completion, you are missing the failure modes that matter most in enterprise settings.

The headline number is 0.663. That is the best score on EnterpriseClawBench, reached by Codex running on GPT-5.5, the strongest configuration tested. The gap between that ceiling and production-ready performance is not a rounding error. It signals that the tasks are genuinely hard, and that hardness comes from realism, not from adversarial construction.

We're thinking: The benchmark data stays private, which is the right call given the internal enterprise content, but it does create a reproducibility ceiling that the community will need to work around. What we find more consequential is the implicit indictment of the current eval ecosystem: if the best available model-harness combination scores 0.663 on tasks drawn from real sessions, then teams shipping enterprise agents today are almost certainly over-estimating their deployment readiness based on GAIA or WebArena scores. The construction and evaluation protocol is the actual contribution here, and organizations with their own session archives can apply it directly. That may be more valuable than any leaderboard number EnterpriseClawBench produces.

Key takeaways:

• EnterpriseClawBench converts real workplace agent sessions into 852 reproducible tasks with recovered fixtures, semantic rubrics, and multi-dimensional evaluation across artifact delivery, visual quality, cost, and runtime, replacing single-score reporting with a profile that reflects actual enterprise failure modes.
• The best tested configuration, Codex with GPT-5.5, reaches 0.663; the benchmark data is not publicly released due to proprietary content, which limits direct external replication but does not affect the reusable construction protocol.
• Teams running enterprise agent evaluations should apply the EnterpriseClawBench construction protocol to their own session archives rather than relying on synthetic benchmarks, and should restructure eval reporting to include artifact quality, cost per task, and skill-transfer behavior alongside completion rate.

Source: EnterpriseClawBench: Benchmarking Agents from Real Workplace Sessions