Stardock CEO's agent benchmark run: only 2 of 5 tasks pass, agents not ready for professional work
draginol · x · 2026-10-19
Stardock founder Brad Wardell shared an agent benchmark run: 2 of 5 tasks passed, average score 0.63 across 38 steps. Failures included a margin-decline analysis (exceeded the hard action limit of 10 steps), a JS zero-value serialization coding task (hit the limit at 11), and SaaS churn recovery (0.79). His verdict: agents aren't quite there yet for professional-grade work — with a hat tip to professional reviewers who do this routinely.
More from coding & agent
- Developer ditches Claude/Codex subscriptions for Bedrock-served coding agents — NYCounihan · 2026-10-19
- Claude and Codex blame each other for bugs, then 'fix' them — only 72% actually fixed — jkubicki · 2026-10-19
- Developer's Codex usage drops to zero since Opus 5.5 shipped — rudrank · 2026-10-19
- Heavy Codex user: Claude's agent harness feels maintained, Codex is 'a big vibe' — karmay007 · 2026-10-19
- Melony: a free frontend/auth layer for Claude Managed Agents, built almost entirely by Claude — DarasStayHome · 2026-10-19
- Claude uses Hedra MCP to auto-refit a music video into 9:16 vertical — AIandDesign · 2026-10-19