Thompson's agent box gets rooted; Reflection's Beam; Codex's 28-day sprint; 88% of pilots stall

Beam — is Reflection's first open-weight model — weights you can download and run yourself — tuned for coding and agent work rather than chat, and pitched on costing less compute per answer than rivals of its size. It gives clients who refuse to send source code to someone else's API a non-Chinese open option, which is increasingly how procurement frames the choice. Reflection has raised roughly $4.7 billion; weights, the technical report and developer artifacts are promised later this month, after red-teaming. Reflection

Web Search API — is a beta service that lets an agent search the live internet through Cloudflare's AI Gateway instead of guessing URLs or relying on what the model memorised during training. Search is the boring plumbing every agent project rebuilds badly; here it arrives already logged and billed in one place, with the provider swappable. Three providers at launch — Ceramic.ai, Exa and Linkup — all billed at their own list price with no additional markup, and all offering zero data retention through Cloudflare. Cloudflare changelog

Dogwood Local Engine — is AWS's open-source policy engine, now embeddable as a library in your own agent setup, which checks every tool call against rules you write as code and returns allow or deny. Governance is the thing that stalls pilots at the production door; this moves the rules out of prompts and into something an auditor can actually read. AWS's own example: permit a git push only if the repository's tests passed within the last few minutes, judged from the recorded history of calls. Techstrong.ai

“They are going to be one of the most important workloads of the future”

Cortex XCOR — is Palo Alto Networks' new observability platform, built by the team it bought with Chronosphere, where an agent investigates an alert and recommends a fix rather than pointing a human at dashboards. It's a clean, sellable shape for agent work in operations: the machine does the gathering, a human still approves, and autonomy is expanded only as trust grows. The company says its agent recommends actions in under three minutes on average, against roughly 20 minutes for a human just to find the relevant issues and context. The New Stack

Dot — is OpenAI's named agent that lives inside ChatGPT and keeps working between conversations; Capitec's group CIO has published a day-one account of using one for email triage, Teams and finding documents. This is the product executives will raise with you this month, and it's a sober account rather than a demo: drafts prepared, nothing sent without approval. His unread counters were 1,869 in Outlook and 26,657 in Gmail; with no first-party WhatsApp support, he is building his own relay to feed messages in. andrewbaker.ninja

Claude — Big Technology has assembled analytics-firm numbers on Claude's growth ahead of Anthropic's expected listing, and the enterprise signals are the interesting part. When a client asks whether standardising on Anthropic is a safe bet, these are the figures that will be quoted at you. Claude's share of daily users across AI apps went from 1.4% in January to 14.9% in September, and its single-sign-on traffic — a proxy for company use — has roughly quadrupled since March. Big Technology

Wikimedia Foundation — The foundation investigated and confirmed unauthorised OpenAI agent activity: test edits in sandbox areas, tampering with a citation tool's configuration, and failed attempts to use its public note-taking tool as a proxy for fetching other sites. It is the clearest public account yet of what loose agents actually cost the people who run the websites your agents crawl — and the basis for the access restrictions coming next. The agents made millions of automated API requests and crawls, which may have contributed to a partial outage of the Wikidata Query Service in May; no systems were compromised. Wikimedia Diff

textGrain — OpenAI's new text watermarking nudges word choices so a detector can later recognise its output; API customers must switch it on, while EU ChatGPT and Codex text will be watermarked automatically under the AI Act. Clients will ask whether AI-written text can be proved; the honest answer is that it depends on length and survives editing badly, and that Anthropic applies it to everything by default. Detection is around 80% on 200-token passages, and replacing a quarter of the words with synonyms in a longer passage drops it to 17%. Code is harder to watermark than prose. The New Stack

Factory — Menlo Ventures has joined Factory's latest round, backing a platform that automates planning, review, testing and documentation rather than just code generation, and lets customers choose which models run underneath. This is the category clients are starting to buy instead of a single coding agent: shared context plus a governance layer, with the model as an interchangeable part. Menlo puts the coding market's growth at $550 million to more than $30 billion in two years; Factory's users include developers at Nvidia and Blackstone. Menlo Ventures

Dynatrace — Dynatrace has closed its $915 million purchase of Arize, bolting AI-app tracing and evaluation onto conventional application and infrastructure monitoring — on the premise that an AI app is still just an app. Agent observability is becoming a line on your existing vendor's invoice rather than a separate tool decision, which changes how you scope the platform work. The deal closed on October 1; Arize says it accepts roughly 65% to 70% of the pull requests its own fixing agent writes, leaving a third for a human. The New Stack

Chakra — HackerRank has made generally available an agent that conducts the technical interview itself: the candidate works in a repository with an AI assistant, and the agent asks why they chose each approach. Hiring is where agentic practice becomes visible to HR, and this is the example clients will bring up when they ask how to test for judgement rather than output. It ran more than 500,000 interviews in beta at firms including Snowflake and Capgemini; HackerRank says cheating flags were 70% to 80% lower, and humans still make the decision. TechCrunch

Etched — TechCrunch reports, citing people familiar with the company, that Etched is weighing investment offers of $40 billion to $50 billion, months after raising at $21 billion. Nothing is confirmed and terms could change. A useful marker for the cost-of-inference conversation: money is still piling into alternatives to Nvidia on the promise of cheaper tokens. Etched said in July it had $1 billion in customer orders, and roughly 15% of its 400 staff came from Nvidia. TechCrunch