AGENTS · 122 POSTS · PAGE 3/13
Agents.
Writing tagged Agents: 122 posts on AI systems, engineering tradeoffs, and building products that have to work in production.
PAUSE intercepts a coding agent's first source edit and injects one plan, lifting SWE-bench Verified by up to 16.8 points at about a tenth of a periodic overseer's Planner cost.
Coding agents followed an outdated instruction file in 90.1% of choices even when they flagged the conflict; one user-level sentence flipped it to 91.4% following the code.
RegretBench shows coding agents rarely back up before destructive steps, and capability gains don't fix it. A cheap interception layer, BackUpper, restored 92.2% of state at $0.014 per task.
A self-improving agent platform's LLM gate silently ignored its input, and regenerated rollbacks failed to restore behaviour. Two cheap checks catch both before you let changes ship unattended.
SWE-Residual shows five frontier agents lose 20.6–53.5 points when they inherit a predecessor's failed attempt, mostly by trusting it. If your agents resume work, test for that.
A pre-registered study of 465 agent-authored PRs found healthier code cost 3.5–6.1% fewer review tokens per point, and file size absorbed the effect. Smaller files are the cheap lever.
Planting the same misconception in an LLM's plan instead of its hypothesis raised repair failure by 29–32 points. Plans are commitments; review and regenerate them accordingly.
Selecting a few files and showing them verbatim resolved 43.7–46.7% of 300 issues, about 6 points above the best compressor, because compressed code breaks the exact quotes patches need.
Re-scoring 600 public agent trajectories, turn count explains 84% of log cost variance and observation size 4%. Cache-aware pricing makes flat-rate estimates ~7x too high.
Four popular efficiency methods were tested on 200 vulnerability-discovery agent traces; only 24.4% of pairs cut cost without losing success. Count helper-model costs before you celebrate savings.