Slok says the AI payoff isn't there, ARC Prize rejects 99.9%, Agent Skills goes to the Linux Foundation

Continuum — Amazon's bug-hunting agent is now generally available: it searches a repository for a genuine flaw, weighs it against your AWS setup, and raises a pull request carrying a patch. Clients will quote the 89% headline at you this month, so it helps to know exactly what that number measures before you agree with it. It cleared 819 of 920 tasks, but only 37.8% of its patches fixed the specific historical bug the test had in mind, and AWS graded the run itself. The New Stack

Data Agent Kit — A free, open-source bundle that connects coding agents to more than fifteen Google data services, shipping both tool access and written best-practice guidance for them. It cleanly separates the two kinds of guardrail clients keep confusing: advice an agent is asked to follow, and permissions it simply cannot get around. One skill opens with a stop-and-verify instruction before anything irreversible such as a DROP TABLE, while an IAM deny rule can block any tool that is not read-only. InfoQ

Enzyme — A small tool that indexes a folder of Markdown notes so Claude Code or Codex can find months-old writing and cite the passage it used. A cheap way to hand an agent the accumulated context of a team without shipping the whole wiki to someone else's cloud. Searching happens entirely on your own machine; Apache 2.0, 98 stars, and the generated question-prompts it builds are explicitly labelled hints rather than evidence. GitHub

Argus — A free, self-hosted security operations platform in which AI agents triage and investigate incidents but are structurally unable to execute any change themselves. A handy reference design when a client asks what agents in a security team should look like: read-only investigation, one broker for changes, everything reversible. Seven specialist agents can only file a request; 27 typed actions pass through policy, dry run and rollback, and both autonomous response and autonomous pentesting arrive switched off. GitHub

Dessa — Jack Dorsey's Block has shown a conversational voice model trained purely on recordings of people talking, with no text and no language model underneath it. Voice agents are the next thing clients ask about, and this one attacks the usual pipeline's weak spot: turning speech into text throws tone, timing and emotion away. It listens and talks simultaneously, was trained on a single NVIDIA GB200 NVL72 rack, and Block says it runs locally on a MacBook Pro M4 Max. Waitlist only, no weights or API. Block

Agent Skills — Anthropic has handed the skills format — a folder with a SKILL.md file telling an agent how to perform one task — to the Agentic AI Foundation at the Linux Foundation, including the GitHub organisation and the agentskills.io domain. Same manoeuvre it pulled with MCP, and it makes the format far easier for a client to standardise on without betting the roadmap on one vendor. Anthropic counts over 70 platforms and a million published skills, yet the standard still has no signing or sandboxing; a Snyk audit in February flagged 76 malicious skills. Simon Willison

Gemini 4 Argon — Google ran its model unattended for eleven hours against the Agent Development Kit, a 200,000-line Python project, with one instruction: cut the overhead without changing the public API or the tests. A rare, honest write-up of an agent working on an old codebase rather than an empty folder — and of the human review bill that arrives afterwards. Two optimisations were merged and four were turned down by reviewers over reliability and maintainability; the agent ran all 16,892 unit tests before it stopped. InfoQ

Chroma — The vector database company has signed to join Replit and is winding down its hosted service over 60 days; the open-source ChromaDB continues under Apache 2.0, with extra code moved out of private repositories. If a client's retrieval or agent-memory layer sits on Chroma Cloud, that is now a dated migration on someone's calendar rather than a someday problem. Chroma reports roughly 6 million downloads a month and nearly 30,000 GitHub stars; named customers include Retool, Paramount and CoreWeave. explainx.ai

Bankrbot — In May an attacker sent Grok's trading companion an NFT with a Morse code message hidden inside; Grok translated it, passed it along as a normal request, and about $200,000 of tokens left for a wallet that should never have received them. The cleanest example you can carry into a client meeting of an agent holding more standing permission than anybody intended, now that brokerages are wiring chat straight to order entry. Musk told a user he would cover losses; the consumer terms reportedly cap xAI's liability at $100, a figure attributed to Cybernews and not confirmed by the company. Startup Fortune

“Try it out. If Grok Bot messes up, we will make you whole.”

Anthropic's field team says one agent often beats many. Anthropic's Applied AI group compared enterprise deployments and came out against fleets of specialised helpers: fold them back into a single loop and give that one agent skills instead, which it found cheaper and faster as well as better. Say this out loud to any client drawing an org chart of agents. Complexity is not the deliverable, and the vendor that invented subagents is now arguing against overusing them. The same roundup covers a test spanning 40 small Python projects where the instruction file and the existing code deliberately disagreed, and the agents trusted the file. Dark Factory Dev

“a single agent with skills frequently outperformed both the one-prompt-for-everything design and the subagent design on quality,”

The crowded market for agents that live in your texts. TechCrunch counted more than twenty assistants you reach by SMS or iMessage instead of installing anything: family schedulers, travel bookers, general chiefs of staff, several with their own cloud machine. This is where consumer expectations are being set — the agent turns up in the thread people already use — and your client's customers will soon expect the same of them. Instinct is valued at $10 billion after a $1 billion round, and Poke was approved in June 2026 as the first AI agent on Apple's Messages for Business platform. TechCrunch

LiteLLM — LiteLLM disclosed that a master key for its Agent Platform was accidentally committed to a public GitHub repository, which an attacker used to publish a malicious version of the Python package and harvest GitHub tokens. LiteLLM sits quietly inside a lot of client model gateways, and this is the second supply-chain scare around it this month — worth an inventory check and a key rotation. The configuration file carrying the key went into the public repository on 27 May 2026, with disclosure following on Friday. The Hacker News

x402 SDKs — Security firm V12 reported that agents paying for API calls through the x402 payment kits could be induced to sign away their whole balance of a token, because the default client carried no spending limit. Every pilot that gives an agent a wallet needs a per-call cap written in; this is the concrete failure that the comfortable default allowed. V12 reported it on 22 October, though releases back on 18 August had already added a limit; only agents signing payments automatically were exposed. The Hacker News