Researchagentsbenchmarksreliabilityopen source
OpenClaw Reveals Agent Reliability Failures In Real-World Tasks
8.2
OpenClaw, a new open-source benchmark released in 2025, tests AI agents on realistic computer-use tasks and finds leading models from OpenAI, Anthropic, and Google fail frequently and unpredictably. Failures include destructive file operations, looping behaviors, and unrecoverable errors, suggesting enterprises should retain human oversight and adopt realistic evaluation before deploying autonomous agents.
Key Points
- 1Shows leading agents complete complex computer-use tasks successfully only a fraction of the time
- 2Highlights failure modes are stochastic, unrecoverable, and potentially destructive, undermining trust for deployment
- 3Implies enterprises must keep humans in loop, add rollbacks, and evaluate agents with realistic benchmarks
Scoring Rationale
Strong industry-wide relevance and actionable findings justify a high score; limited peer review and single-source reporting reduce certainty.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems