AGENTS · 122 POSTS
Agents.
Writing tagged Agents: 122 posts on AI systems, engineering tradeoffs, and building products that have to work in production.
A placebo-controlled study of 5,835 coding-agent runs found telling the agent its token budget never beat an unrelated sentence. The cap itself is where the leverage is.
HerAgent's ablation shows that environment-setup fixes made in a live container but never written back into the setup script mostly fail a clean rebuild: 25.9% vs 90.7% final success.
Asking agents to review their work barely helped on DeepSWE. Framing the same check as advice to the spec's maintainers lifted pass@1 from 44.25% to 50.22% at under half Self-Refine's cost.
Across 6,541 agent commits, 37% of checkable AGENTS.md rules were never followed and 54% always were. Rewording rarely helps; give each rule a check or a hook.
Compiling one recorded Claude Code skill run into programs cut repeat-run tokens 74–81% on 87 SkillsBench tasks; compiling from the skill text alone made most tasks costlier.
FlowPlan makes Claude Code state its fix as a control-flow plan and check the edit against it, cutting tokens to as little as 0.32x with no significant change in resolution rate.
BoxBox let a coding agent edit its own harness during a run and reached 1.702× speedup at lower cost, while offline harness evolution scored below the frozen baseline.
Stale AGENTS.md lines barely changed task success across 732 runs, but a dead test command produced silent no-op checks in 64 of 80 runs and no agent ever fixed the file.
Trajectory-based failure predictors for coding agents added nothing over issue difficulty plus model identity, and model switching gains were within noise. Test your alarms against run-blind baselines.
In Multi-step SWE, only 12% of failed steps used their full time, and 52% ended with the agent claiming completion. Treat "done" as a trigger for independent checks, not a verdict.