// PAPERS
One paper a week,
engineer-first.
New arXiv work on agents, tool use, long context, and coding agents — read from a builder's angle, in four languages, every Saturday.
Beyond the Leaderboard: The 6 Failure Modes That Kill Real Agents
New arXiv work (Rao et al., 2026) meta-analyzes 27 agent benchmarks and 4,300 failed trajectories. Six failure modes explain 89% of them — and only two are what the leaderboards actually measure. A field guide for anyone shipping Claude Code in production.
FastContext: When Repo Exploration Deserves Its Own Sub-Agent
New arXiv work (Kim et al., 2026) argues repository exploration is a distinct capability from patch generation — split them and you get +5.5 points on SWE-Bench with 60% fewer tokens. A deep read for anyone building Claude-Code-style loops.
Self-GC: Teaching Agents to Fold Their Own Context
New arXiv work (Sun et al., 2026) turns agent context into first-class objects the model can fold, mask, and prune. Result: 3–5× longer trajectories at the same token budget, and a cleaner mental model for anyone building Claude-Code-style loops.
Coding Agents Beat Long Context: What the Duke Paper Actually Shows
New arXiv work (Cao et al., 2026) shows off-the-shelf coding agents outperform SOTA long-context systems by 17.3% on average across five benchmarks. Here's what changed, why it works, and what to steal for your own agent.