AI Agents

This is an honest census of the MCP ecosystem: 400 npm/stdio servers drawn with a published seed from a 24,135-server registry, probed over the wire. Only 48.8% complete an initialize handshake versus 66.7% for a hand-curated frame, and the dominant failure isn't missing credentials (13.3%) but servers that never start (37.5%). Among the 195 that run, JSON Schema conformance is total — zero fatal violations across 2,766 tools — but 58.8% omit safety annotations. Curation systematically flatters every reliability number.

A production skill router over 34,396 skills documents a real trap: fine-tuning on synthetic data improves in-distribution retrieval but causes catastrophic forgetting on real and out-of-distribution skills. The fix borrows continual-learning methods — embedding-anchor regularization, LwF, EWC, L2-init — which not only preserve OOD retrieval but also improve synthetic in-distribution retrieval by 13.98% for a 0.6B Qwen retriever and reranker. The scale and the multi-positive, scarce-supervision setup make this more than a toy result.

DriftNet reads a logged tool-call trajectory and answers three operator questions in one forward pass: is the trajectory compromised, where did the injection enter, and which steps were hijacked versus resisted. A dual-head Transformer under two million parameters, using a frozen sentence encoder and identity-free step features, needs no access to the agent's model. On the task-disjoint AgentDrift split (12,536 trajectories), it hits 0.983 trajectory F1, 98.7% exact injection-point recovery, 0.979 hijacked-span IoU, and zero flags on 218 resisted attacks.

EvidenceNet targets a coordination problem in network automation: agents with different authority scopes each make config changes, but a successful local action doesn't prove remote devices responded or that routing reached where it needed to. A runtime assurance layer collects post-change observations required by a completion contract, and an admission gate checks the evidence comes from the right scopes, stays current, and satisfies task rules before declaring the coordinated operation complete. Validated on live-network experiments rather than simulation.

This names a failure mode most eval harnesses hide by averaging: on AppWorld with ReAct/GPT-4.1, per-run pass rate averages 77% but the agent succeeds in all five runs on the same task only 53% of the time — a 24-point consistency gap. The proposed fix is a Consistency Analyzer that finds unstable steps likely to flip across runs, plus a Guideline Generator that writes targeted guidance into episodic memory. Result: +16 points on same-task and +13 on similar-task all-five-runs success.

OpenAI published overview docs for an Agents API, a first-party surface for building agents against their platform. It drew 294 points and 159 comments on Hacker News, so the community reaction is real, but the raw text here is just the landing page — no specifics on tool-calling contracts, session/state handling, or how it relates to the existing Responses and Assistants surfaces. Worth reading the docs directly before forming an opinion on whether it changes your framework choices.

Two Datasette security patches (1.0a39 and 0.65.4) came out of an audit run with Claude Fable 5.1, GPT-5.6, and GPT-6 Astra, followed by nearly a week of human review. The workflow detail is the payoff: for each issue, one person wrote the automated test exposing the bug while the other implemented the fix, keeping two sets of eyes on every change. The models surfaced subtle bugs; Willison says frontier-model audits are now standard practice going forward.

Shopify is moving mobile back from React Native to separate Swift and Kotlin codebases, reversing a 2020 decision. The reasoning is squarely agent-driven: the cost of maintaining two native platforms hasn't gone away, but agents now do enough of the implementation, translation, testing, and review that duplication is no longer the deciding factor. It's a well-argued post that credits React Native fairly, but it's a directional call — no measurements on how much of the two-platform work agents actually absorb.