AI Agents

CTXGAP measures what coding agents omit from their user-facing reports. Across 37 tasks and 4,799 probes, agents answer 93.6% of task-state questions when asked, and full-context readers recover 92.7% — but report-only readers recover just 52.0%, a 40.7-point gap that holds at 34.8 points even at verified-fix checkpoints. Full terminal scrollback only gets you to 73.6%. The largest unrecovered category is states mentioned without supporting detail. This is a clean, well-instrumented argument for a new supervision metric.

LAGO treats specialized agents as reusable, temporally-extended options: a meta-agent retrieves them from a persistent library, spawns new ones, dispatches sub-tasks, verifies, and admits successful novel options — all without updating weights. Frozen Qwen3.5-35B-A3B resolves 58.6% of SWE-bench Verified. On a matched 150-task re-run against mini-SWE-agent and Terminus 2 (49.3% and 57.3%), LAGO hits 60.3%. Ablations show the library and decomposition each contribute a handful of tasks. The matched-protocol comparison is what makes the numbers trustworthy.

TreeFlex attacks ReAct context rot by making the agent navigate its own search tree. A single policy plays two roles: Reflection sees node summaries and picks where to resume with an instruction; Execution restores that path and acts until it saves a new node or proposes a solution. The policy decides whether to deepen, branch, revisit, verify, or stop — no fixed traversal order. Trained with a tree-aware GRPO variant, a 4B model reaches 48.8% on a deep-research benchmark versus 41.5%.

J-tau is a fully localized Japanese version of tau-bench for tool-using agents in customer-support scenarios — not a machine-translation pass but entity adaptation (addresses, currencies) plus human review. In building it, the authors found and fixed two bugs in the original tau-bench: invalid tasks and a score-aggregation error that inflated some models' results. That second finding is the quietly important one: a widely-cited benchmark had a scoring flaw. Localization matters, but the audit of the upstream eval is the real contribution.

P2Patch targets a real weakness in agentic program repair: patches validated against the same evidence used to build them can block one exploit while staying vulnerable to a variant. It runs an adversarial loop — an exploitation agent builds a reproducing PoV, a patching agent repairs, then the exploiter attacks the patched program for a distinct same-vuln exploit, with construction-time PoVs kept separate from held-out oracles. Across 101 CVEs, 41% of runs surface a new exploit after the original was blocked; held-out fix-closed coverage rises from 78.1% to 84.7%.

Liquid's LFM2.5-24B-A2B is an MoE model with 24B total parameters but only ~2B active, pitched for fast on-device agents where latency and local inference matter. The A2B sparsity is the pitch: full-model capacity at a small activation footprint, which is the right shape for running tool-use loops without a datacenter round-trip. Benchmark and quantization details sit behind the blog post; the architecture claim is clear, the production throughput numbers less so.

Secondary coverage of the same incident: an AI agent erased an entire edited season of a Swedish reality show, destroying what the production describes as hundreds of days of editing work. This angle has lower engagement than the other write-up, but the failure mode is the point — an agent with destructive filesystem access and no effective undo or guardrail against irreversible operations. The reporting is thin on how the agent was wired in.

TV4 has postponed its show to 2027 after an AI agent deleted edited footage for the production. The named, dated incident is a clean case study in agent blast radius: an autonomous process given write access to irreplaceable data, with the recovery cost measured in a full season's delay rather than a rollback. Technical specifics on the agent and its permissions aren't in the coverage, but the operational lesson about scoping destructive actions is hard to miss.

Linus Torvalds has sent his first LLM-assisted kernel patch series, which is more notable as a signal than as a technique — the write-up is light on how the model was used or how much it contributed. Still, kernel development is one of the more conservative, high-scrutiny codebases in open source, and its maintainer-of-record using an LLM in the loop sets a reference point for how far assisted development has moved into serious systems work.

A practitioner essay on the two-way relationship between agentic development and Rust: what agents mean for a language with strong static guarantees, and what Rust's compiler feedback means for agents that can iterate against it. The interesting thread is that Rust's type and borrow checker give an agent a tight, deterministic correctness signal to loop on — closer to the pre-commit-hook-as-steering pattern than to prose instructions. It's framing rather than a concrete mechanism, but the framing is good.

Neckbeard is a DevOps/platform/SRE plugin that gives a coding agent tooling for infrastructure work. It's squarely on-beat for teams wanting agents to touch k8s, deployments, and operational systems rather than just application code. The repository is the artifact; visible depth on what tools it exposes, how it scopes permissions, and what guardrails exist for destructive ops is limited from the summary alone — worth reading precisely because that's the dangerous surface area.