AI progress — 23 October 2026

Anthropic is donating the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation, the same body that took over MCP. That moves a format many agent tools already use out of one vendor’s hands. Separately, 1Password announced a way for agents to log in to accounts without seeing the password, and Ars Technica reports attackers were probing the PHP2Shell bug before defenders knew of it.

New models and papers

• Is Complex Agentic Memory Always Needed? — a paper testing a plain RAG baseline (dense retrieval plus reranking over raw conversation history) against agentic memory systems on benchmarks including LoCoMo and LongMemEval. The title states the authors’ conclusion that naive RAG is stronger than assumed. Why it matters: if the result holds on your workload, a simple retrieval baseline is the thing to beat before adding a memory layer.
• LLMs and formal software verification — an Anthropic research post arguing that LLMs make formal verification feasible at a much larger scale. Why it matters: formal proofs have been limited by the labour of writing them; the post’s claim, if borne out, targets that cost. Only the title was available to this digest.
• OlmoEarth v1.3 — Allen AI’s updated embedding model for Earth observation, billed as state of the art for the whole planet. Why it matters: it is a ready-made embedding source for remote-sensing work; verify the benchmark claim against your own data.

New tools and software

• Agent Skills donated to the Agentic AI Foundation — Anthropic is giving the Agent Skills standard to the AAIF under the Linux Foundation, and says it will stay open and vendor-neutral. Why it matters: skills you write for one agent tool now sit under neutral governance, which lowers the risk of building on the format.
• Claude Code v2.1.309 — fixes several sandbox gaps: shell startup files running outside the sandbox on macOS and Linux, symlinked denyWrite globs, and modification of the –mcp-config file by sandboxed commands. Why it matters: if you rely on the sandbox to contain agent commands, these were ways around it.
• SageMaker training session APIs — lets you write your own LLM training loop (SFT, RL or custom objectives) in Python while the service runs the GPUs; the post walks through a GRPO example. Why it matters: custom RL loops no longer require you to operate a GPU cluster.

Overall digest

• 1Password lets agents log in without seeing the password — a new feature lets agents sign in to a customer’s personal accounts while the credentials stay protected. Why it matters: it addresses the usual alternative of pasting secrets into an agent’s context.
• PHP2Shell: attackers had a four-day head start — Ars Technica reports hackers were probing the bug before defenders were told it existed. Why it matters: anyone running PHP should check patch status and look for activity in the pre-disclosure window.
• Meta turns Muse sharing back on — Meta re-enabled Muse sharing after researchers at Claroc reported that public conversation links were easy to find, including pages that looked like internal business documents. Why it matters: shared AI chat links can be discoverable; treat them as public.