AI Agents

Silent tool failures — a cached error page or negative price arriving in the expected format — get consumed as fact because they don't trip a timeout. Outcome Monitors detect violations of outcome contracts mined from task-disjoint traces or public schemas, then issue a nonbinding receipt naming the violated property and recovery tools. ToolMaze completion rises from 10.9% to 28.1% across four models; tau-bench retail gains 12-14 points. Ablations show the recovery-tool list is the active ingredient.

Clearing a rejected branch from the transcript doesn't clear it from the model's attention if the serving session retains KV state across the abort. The paper formalizes rollback consistency and uses a same-token/different-cache audit to isolate the effect: retained KV alone flips a protected effect in 25 of 63 cells across seven open-weight families, with attacker tokens absent from the served request. Reproduces on Hugging Face Transformers cache-reuse and LangGraph time-travel. Rebuilding the cache closes every cell.

A risk-controlled framework for reference-free LLM judging on objective tasks. It calibrates uncertainty thresholds on a held-out set using finite-sample Clopper-Pearson intervals so the false discovery rate among accepted verdicts stays below a user-specified α. When the parametric mode isn't confident enough, it routes to a retrieval-augmented mode with a second calibrated threshold, and the guarantee carries over. This is the missing piece in most judge pipelines: a formal handle on accepted-verdict risk instead of vibes.

Instead of prompting an agent to generate tests directly, this has it first document pre-conditions, post-conditions, and undefined behaviors as a semi-formal spec that scaffolds the test generation. Evaluated on production bugs from Google: 9.8pp higher bug detection (p=0.0352) and 2.5pp better branch coverage (p=0.0034) over a direct baseline. LLM-as-judge rates the spec-driven suites above both baseline and human-authored tests in a majority of cases. Real bugs and reported p-values make this more than a toy result.

Agent Skills package reusable procedures with executable resources, and generating them automatically risks shipping unsafe side effects. TRUSS gates candidates statically against source evidence and nine safety properties, then loads survivors into a shadow agent in a controllable execution environment where brokered tools enforce policy and record provenance traces. On the benchmarks it hits 100% precision and recall in vulnerability detection, cuts attack success from 38.7% to 19.4% with GPT-5.5, and raises task effectiveness from 17.1% to 52.9%.

The official MCP roadmap lays out where the protocol is heading, which matters if you're building tool-use layers on top of it. HN engagement is heavy (216 points, 133 comments) but the newsletter can't verify specifics without the linked post. Treat this as a signal to check what's changing in the spec before it lands in a breaking release, not as a settled description of concrete changes.

An open-source, self-hostable IDE that runs Claude Code, Codex, OpenCode, Cursor, and Grok in one place, with inter-agent delegation and automations for recurring work. The pitch — a horizontal agent app as the main company interface — is founder framing, and the launch is light on the operational detail senior engineers care about (isolation, cost accounting, failure handling). The self-hostable multi-agent angle is the interesting part; evaluate it as code, not as the thesis.

A reported A/B test of reduced effort levels in Claude Code — the kind of silent server-side change that shifts your agent's behavior without a version bump. If real, it's a concrete argument for pinning behavior with your own evals rather than trusting perceived model quality. But this is an unconfirmed Twitter claim with no methodology behind it; the 192-point HN thread is reaction, not verification. Worth watching, not acting on.

A forum writeup arguing that local LLMs underperform their reputation because of quantization, default sampling, and templating mistakes rather than the weights themselves — relevant if you're self-hosting models behind agents. It drew 366 points, so the pain is widely felt. The full text isn't available here to verify how deep the diagnosis goes, but the framing (misconfiguration, not model capacity) is the right one to check against your own deployment.