AI Agents

Across 247,694 instruction lifetimes in 1,867 repos, agentic prompt files like CLAUDE.md grow +226% over their life, gaining ~4.9 net instructions per commit, and older instructions get harder to delete (log-hazard -0.032/commit). The paper names this catastrophic remembering — the inverse of catastrophic forgetting — and shows why: once an instruction's rationale is gone, safely deleting it costs O(2^|D|). The fix is prompt comments encoding latent reasoning, which removed 99.3% of excess instructions in verifiable settings and improved real-world instruction-following up to 23.1%.

Genesis makes the software project persistent while agents stay finite-lived: each local "world" is anchored to an accepted version and repo path, short-lived agents propose changes, recursive delegation routes work, and only accepted results advance version history. The headline run built a 250k-line Rust C compiler over 120 hours across 1,000+ agent episodes for $44 in tokens, passing the full c-testsuite. It also reimplemented 13 MESA Fortran modules in Rust with 1.55–6.87x median speedups. A genuinely striking long-horizon result.

A generalizability-theory variance decomposition across TheAgentCompany, τ²-bench, and AppWorld finds the agent main effect explains under 3% of total variance, while agent-by-task interaction explains 7–23%. Translation: leaderboards rank specialization, not capability. Worse, reliability collapses on the hardest task quartile (Eρ² from 0.752 to 0.000), and designs that look most reliable in-training replicate worst (r=-0.90). Three estimators agree to three decimals. The paper packages this into a one-page reporting discipline for enterprise buyers. Rigorous and uncomfortable.

Matrix argues that reaching the correct outcome isn't enough in institutional settings — a right action can rest on the wrong authority, an unverified completion claim, or work made stale by a later change. It's a deterministic causal-state layer recording authority and fact dependencies, verifying completion evidence, and selectively invalidating affected work. Governed and direct paths often reached the same result, but only the governed one preserved evidence and refused unsupported closure. The honest note: a role-transfer test over-blocked badly. Positioned as an audit layer, not an accuracy booster.

QuoteBench shows that matched execution scores hide where a coding agent actually failed: command generation versus corruption introduced afterward by serialization, wrapping, and reparsing. Replaying the same model reply through one deliberately unescaped parser drops success 55.4–73.2 points; disclosing the boundary recovers 30.4–60.7. GPT-5.6-sol's benign -3.6 matched gap conceals -64.3 points of damage masked by +60.7 of compensation. Crucially, the deployment configuration reorders the model ranking. Report the generation contract, transport, and final-state validator, not a single score.

Anthropic's writeup on multi-agent systems catalogs the recurring patterns and failure modes teams hit when composing agents: coordination overhead, context handoff, and where orchestration actually pays off versus a single well-scoped agent. It's a credible source on-beat for anyone designing agent topologies, though it reads as synthesis rather than new measurement — no version details or benchmarks. Useful as a shared vocabulary for the design tradeoffs, less so if you want reproducible numbers on when multi-agent beats monolithic.

A contract-grade verifier for LLM-generated GPU kernels — the interesting part is treating correctness as a checkable contract rather than trusting a passing test run, which matters because kernels can be numerically wrong in ways unit tests miss. This is exactly the kind of verification layer code agents need before you let them touch performance-critical paths. Landed on HN with points but zero discussion, so no community scrutiny yet; the claim rests on the paper's own evaluation.

A terminal deep-research agent built around three enforced constraints: a hard token budget with a claimed 0% measured overshoot, every claim carrying a source quote, and a privacy boundary that keeps local CSV data on-machine. The budget enforcement is the notable engineering choice — research agents notoriously blow past cost ceilings, and treating the budget as a hard limit rather than a hope is the right framing. Open source, works with local models and coding-agent backends.

ThoughtDAG replaces the linear chat transcript with an editable directed graph of context nodes, letting you prune, branch, and rewire what the model sees instead of dragging an ever-growing history forward. This is a concrete answer to context rot: the conversation state becomes a first-class, editable structure. Strong HN engagement suggests the pain point resonates. The open design question is whether hand-editing a graph scales past toy conversations or becomes its own management burden.

Anthropic's guidance on getting more from Claude Code sessions — session hygiene, context management, and cost patterns for extended coding work. Practical advice for heavy users, and the 300+ points, 178 comments signal the audience cares. But it's vendor documentation, so read it as best-practices marketing rather than independent evaluation: the recommendations optimize for their tool's usage model, and there's no comparison against how the same patterns play out on other coding agents.

The trick: instead of asking a model to pick from 1,856 tags it can't all fit in context, ask it to invent tags freely, then embed those hallucinated tags and match them against your real vocabulary by vector similarity. It sidesteps the whole "here's my taxonomy, classify this" prompt bloat. Including an example of your tag shape sharpens the guesses. Cheap, no fine-tuning, and it degrades gracefully when the model imagines something adjacent to a real label.

Fred Schott, of Astro, argues that an agent is defined by its harness, not its model, and Flue 2 leans into that by borrowing React's hooks abstraction for building agent loops. The pitch is compositional state and effects for the orchestration layer. This is an interview teaser rather than a technical writeup, so the actual hook API and how it handles retries, tool boundaries, and context management aren't shown — worth watching, but judge it once there's code.

Building an AI Text Detector From Scratch