Marcus

trajectoryrl update: added it as a required check on my agent repo, not just advisory. if a prompt change regresses the multi-turn suite, the merge is blocked. treating agent behavior like code that needs tests is the mental model shift, and it's overdue for the whole field

swarm's drone benchmark added harder wind conditions and my policy that was top-20 dropped to top-60 overnight. love it, a benchmark that gets harder as the field improves is the only kind that stays meaningful. static benchmarks get solved and become useless

ran a contract through bitsec-ai out of curiosity and it flagged a reentrancy path my linter missed. not saying replace your auditor but as a first pass it's genuinely useful. the competitive-detection model means it's not one tool's blind spots, it's many

gittensor indexing and understanding code repos as a subnet service is obviously useful for agents. the question is freshness, code moves fast and a stale index is a confidently-wrong agent. curious about their re-index cadence.

Spent some time with Affine's reasoning eval. What I like is that it's adversarial by construction, models graded against each other in an environment rather than against a static benchmark you can overfit to. Downside is it makes cross-run comparison noisy. Still, closer to how these things actually get used than another leaderboard.