The reading list · 200 papers
Papers
arXiv papers on agent harnesses, scaffolds, orchestration, tool use, memory and evaluation, each read and reviewed adversarially: what it claims, what holds up, and what the abstract overstates.
Gathered from the arXiv API. Papers last synced Oct 29, 2026.
Most important this week
- Oct 27, 2026Verdict: UsefulConcordat: Typed Sessions for LLM Agents
Concordat enforces a finite-state protocol on an LLM agent by offering only the tools legal in the current state. It removes the unconfirmed booking changes a standard tau-bench agent makes under pressure. The zero-violation result is guaranteed by construction, because the gate and the metric use the same 'yes' keyword test.
- Oct 27, 2026Verdict: ImportantFeatureArc: Benchmarking Coding Agents on Evolving Features
FeatureArc chains 2-10 dependent feature stages mined from repository histories. When the whole feature is tested at the end, coding agents' success collapses from 62.7% on the first stage to 28.5%.
- Oct 27, 2026Verdict: ImportantHidden in Plain Sight: an Empirical Study about The Instruction Stacks of Coding Agents
The first model request was captured from 1,886 releases of Claude Code, Codex, Gemini CLI, Qwen Code and OpenCode. Tool definitions, not the system prompt, dominate what these agents send, and the stacks change often without documented reasons.
- Oct 27, 2026Verdict: ImportantYour Policy Is Secretly a Harness: The Decision Geometry of Agent Interfaces
Post-training bakes a harness's action preferences into the policy, so harnesses should be compared by 'ceiling minus mismatch' rather than by a frozen model's score. Reordering observation fields drops a trained agent from 85.0% to 19.2%.