AI Agents

SpecPath names a failure mode called specification-path sensitivity: coding agents that resolve a final contract correctly when stated directly produce behaviorally different programs when the same contract arrives via an equivalent revision history. Holding repo, contract, verifier, agent, and budget fixed and varying only the path, aggregate accuracy barely moves — yet 35 of 100 blocks that pass on the direct spec fail on at least one equivalent history. The lesson: agents follow the most salient statement, not the active contract, and single-shot benchmarks hide it.

SWE-Bench ProMax targets code refactoring — coordinated, behavior-preserving changes across many files — as a harder replacement for saturating benchmarks. 170 expert-curated instances across seven languages, averaging 11.4 modified files and 261.6 lines each. The curation directly addresses the audit finding that ~60% of unsolved SWE-bench Verified instances have flawed tests: issue descriptions rewritten from scratch, test suites manually reviewed to drop overly narrow and overly broad checks. The scale and the deliberate attack on prior benchmark rot make this a credible successor.

OpenCodeReview attacks two problems in LLM code-review agents — non-determinism from unbounded tool use and context bounded to the diff — by injecting determinism at three points: rule-guided file/criteria dispatch, a curated ReAct tool set with parallel per-file subagents, and an independent reflector that sees only the diff to filter hallucinated comments without self-reinforcement. On AACR-Bench (200 real PRs, 10 languages, 1,505 verified comments) it beats mainstream coding agents on SEM-F1 with 5-15x token savings, and it's open-sourced.

Agent Memory Distillation is a training-free framework that transfers structure from a large teacher agent to a small student via three memory types: workflow memory for task strategy and subtask memory for behavioral examples, both injected proactively, plus function memory retrieved reactively on tool-calling errors. With GPT-5-mini as teacher and 4B-8B students, it adds 27.2pp on AppWorld, 11.2pp on BFCL V3, and 3.4pp on ToolSandbox. Subtask memory drives most of the gain, and 4B students benefit most — a clean, well-ablated result.

RETRACE adds an independent verification signal for coding-agent patches, filling the gap where self-refinement just re-judges a patch under the same interpretation that produced it. It reconstructs forward — building a repair rationale from issue and trajectory — then backward, inferring from the patch alone (no access to the original issue) what problem it appears to solve, and compares the two for an alignment verdict before a reconciliation stage submits or produces revision guidance. Training-free, evaluated on SWE-bench Verified with two backbones and reported Pass@1 gains plus ablations.

Needle 2 is a 45M-parameter agentic model at 2-bit compression, shipping as a single 14MB binary that runs a full session in 28MB of RAM. It decodes 500 tok/s on a Raspberry Pi 5, 400-1,500 on Quest 3S and Vision Pro, and 300-700 on sub-$200 phones. On tool-call and device-use benchmarks it trades wins with LFM2.5 230M and Apple's Foundation Model at 5-70x smaller. The architecture is Simple Attention Networks, with a confidence-based cloud-escalation pattern for the hard cases.

An analysis of Claude and GPT knowledge cutoffs versus their actual pre-training timelines, useful for anyone reasoning about why an agent confidently references a library version that doesn't exist or misses one that does. The post reconstructs the gap between stated cutoffs and observed behavior. The HN discussion is thin (21 comments) and the body is behind the link, so treat the specific dating conclusions as the author's inference rather than confirmed fact.

Ante is a coding agent packaged as a single binary that runs fully offline — directly on-beat for anyone who wants agent-assisted work without shipping their repo to an API. The HN post drew 135 points and 80 comments, but the provided text carries no engineering detail: no model backend, context handling, tool interface, or how the offline constraint shapes capability. The pitch is clear; the substance is behind the repo.

A tiny LLM running at 21,000 tok/s on a $250 Xilinx KV260 FPGA, with a live demo. The throughput number is the headline and it's a genuine hardware result rather than a simulation — worth a look if you're thinking about deterministic, fixed-cost inference at the edge. The agent angle is weak: no tool use or planning, just raw decode speed on a small model. Details on model size and quantization live in the linked writeup.

Dan Luu on which programming language is best for coding agents, framed largely through token efficiency — how compactly a language expresses the same logic and what that costs across an agent's iteration loop. It's a blog-style comparison, not a controlled study, but Luu's analyses tend to be careful and the token-cost angle matters directly for latency and spend when an agent rewrites a file dozens of times. Pair it with the related January piece on token efficiency across languages.

Meta is back in open weights with Muse Glimmer, a 30B model under a clean Apache 2.0 license and pitched squarely at agentic work: end-to-end task completion on SWE-Bench and τ-Bench, precise tool-call schemas, and long-horizon planning. Simon Willison ran the 18GB LM Studio build locally; the pelican SVG came out jumbled but recognizable. The vendor benchmark claims are the vendor's, but a permissively licensed 30B tuned for tool use and multi-step reasoning is worth pulling down and testing yourself.