Dirac tops TerminalBench on Gemini 3 Flash

Dirac’s result is a reminder that benchmark outcomes are often as much about harness quality as model choice. If the run holds up, it strengthens the case that context curation, edit precision, and tool orchestration can swing agent performance materially.

• –The reported 65.2% TerminalBench 2.0 score would put an open-source agent ahead of both Google’s own submission and the current closed-source leader cited in the post.
• –The author explicitly says no `agents/skills.md` files were inserted, no resource or timeout changes were made, and the exact GitHub codebase was used for the run.
• –Dirac’s positioning around hash-anchored edits, AST-aware manipulation, and token efficiency fits the kind of workflow TerminalBench is meant to stress.
• –The post also highlights a real benchmark problem: if the community doubts compliance, the score matters less than the reproducibility story around it.
• –Until the leaderboard accepts the submission, this reads as a strong but still provisional signal that agent scaffolding can be a competitive advantage.