Claude Message Threads | AI Field Notes #130

AnalysisA coding agent deep in a 900,000-token session used to send that whole history back to Anthropic every turn. Message Threads, a Messages API feature Anthropic put into public beta on November 2, sends only the new messages plus the ID of the last response, and keeps the thread in encrypted memory for five minutes between calls. For requests of 500,000 to 1 million tokens, median time to first token (the wait before a reply starts) falls about 35%. The bill stays put, since Claude still reads the full context. Claude Code has it on by default, so most of its users got faster without touching a setting.

AnalysisTwo sentences added to Claude Code's system prompt for Opus 5.5 now stop the agent from recreating existing video games, Anthropic's Addy Osmani confirmed on X, as The Decoder reported on November 3. The rule had been live for several days and holds even when a user renames the clone or calls it original. It follows hundreds of unofficial browser ports of titles like Halo and GTA Vice City. Original games, genre homages and projects with custom assets are still allowed, though some developers report refusals on their own work. Osmani says the rules will keep shifting, which puts the boundary inside a prompt the user cannot read.

AnalysisCopilot now sits as one option in JetBrains Air's agent picker, next to whatever else a team already runs there. The November 3 integration rides on ACP, the Agent Client Protocol (an open standard that lets any coding agent plug into any editor), so JetBrains needed no custom bridge. Developers can run several Copilot sessions in parallel, locally through an alpha IDE plugin or as cloud tasks in isolated environments. Cloud work bills against the Copilot plan instead of JetBrains AI credits, and Air's own governance and analytics cannot see Copilot usage. Admins can switch Copilot off, but they cannot cap its spend from inside Air.

AnalysisEvery Apple developer got access to Siri AI on November 3, months before customers see any third-party integration. The developer preview ships in the iOS, iPadOS, macOS and visionOS 27.3 betas and works through App Intents (the framework apps use to expose actions like booking a table). A new Skills layer lets Siri chain several of those actions for one request, and Siri can now surface anything in the Spotlight semantic index instead of only items that fit Apple's preset schemas. App Store submissions open in the spring. Server-side features carry daily limits that may later cost money, and none of it works in the EU on iOS.

AnalysisIn July, Goldman's CIO Marco Argenti was praising non-technical staff for vibe coding (building software by describing it to an AI). By November 2, Bloomberg reported, Goldman had restricted traders and bankers from advanced coding tools such as Claude Code and GitHub Copilot, leaving access with engineers. Some traders had spent months using the tools to monitor trading flows. A spokesperson cited "productivity and risk management" and said the experimentation phase is over. The plan where a trader builds a prototype and hands it to an engineer to harden is gone, and the bank has not said what replaces it.

AnalysisA researcher who finds a critical hole in ChatGPT or Codex can now collect up to $30,000, or up to $100,000 for an exceptional find, about four times the old rates, OpenAI said on November 3. Stack the new quality, loyalty and campaign bonuses on top and a single report can reach roughly $172,000. Anyone who reported a bug in the last 30 days gets the difference paid retroactively. OpenAI is also lining up its next internal security sprint with the firm Cathedral; in the last one, a quarter of its product engineers spent the sprint hunting bugs. Apple and Microsoft still pay more for operating-system exploits.

AnalysisProving that an API answer came from the model you paid for costs 70 to 95 times the GPU time of running it. That is the honest number in FlashProve, an open-source prototype Anthropic published on November 2. A lab publishes a commitment (a cryptographic fingerprint of the exact weights), and each response can then carry a zero-knowledge proof, math that confirms the claim without exposing the weights. On one H100 the prover manages about 110 tokens a second for Llama 4 Maverick, each 1,024-token proof weighs 19 MB, and an ordinary CPU checks it in about 17 seconds. No lab plans to deploy it yet.

AnalysisCrowdStrike's new off button for AI agents works by starving them. Launched on November 3, it blocks a misbehaving agent's route to the models that drive its decisions through connected AI gateways, keeps the agent from restarting, and pulls it off the corporate network, while leaving its code in place. A human still decides when to pull the switch and when to restore access. The feature lands in Falcon Guardian, the agent-security product CrowdStrike launched in September, alongside planned acquisitions of Fiddler AI (agent observability) and Dreadnode (automated AI security testing) plus an integration with Anthropic's Claude Cowork. No price or availability date came with it.

AnalysisUS export controls stop Nvidia shipping its top chips to China, yet they mostly leave a Chinese lab free to rent the same chips from a data center abroad. Nvidia is now nudging that gap shut on its own. The Information reported on November 3 that Nvidia has asked cloud providers, including neoclouds (smaller GPU-rental clouds, some of which it backs), to take less business from Chinese companies and to disclose the end users of its chips. The request was passed on verbally, is voluntary and covers new contracts only. It follows US scrutiny of overseas compute rented by Moonshot AI, the developer of the Kimi chatbot.

AnalysisCharacter.AI, Replika and any app whose virtual companion engages in sexual or romantic content will carry a minimum 18+ rating on the App Store, under rules Apple sent developers on November 2. A new Dating and Relationships descriptor appears in App Store Connect on November 16, and every app must complete it by January 12, 2027. The same descriptor covers human dating apps, while a separate Shared Location question flags apps that broadcast where a user is. Top AI companion apps had about 220 million downloads by summer 2025. The rule reaches general-purpose apps too, if a romantic companion is only one feature among many.

AnalysisServing Grok Bot, SpaceX's personal AI agent, was "unsustainably high relative to revenue," Elon Musk wrote on November 1, after users found their weekly limits maxed out over Halloween weekend, some jumping from 50% to 100%. Musk confirmed the throttle was deliberate. At a reporters' dinner reported on November 3, OpenAI's Adam Fry called its Dots agents "on the compute heavy side"; each Dot runs on its own cloud computer, and only $100-a-month Pro subscribers get them. OpenAI's fix is a router that sends some tasks to the smaller GPT-6.1 Terra instead of GPT-6 Astra. Always-on agents have run into the price of always-on compute.

AnalysisChanging 0.016% of a model's weights by reasoning about them, with no training run at all, took Qwen3.5-0.8B (a small open model from Alibaba) from 7.9% to 48.1% on the MATH-500 benchmark. MIT CSAIL researchers including Omar Khattab published the result on November 2 as Weight-Space Programming: OpenAI's GPT-6 Astra read the raw weights, found what made the model loop endlessly, and subtracted one direction from 64 columns in two layers. Instruction-following on IFEval rose from 21% to 62.3%. Astra also wrote a Breakout-playing network from scratch that outscored conventionally trained ones. The authors call it very early, and a 270M-parameter Gemma model did not improve.

AnalysisSignal analysis that took weeks now finishes in hours at the SETI Institute's Allen Telescope Array in Northern California, where AI agents steer the telescope and read incoming radio data as it arrives, Nvidia said on October 30. Researchers still type the commands, and the agents handle pointing and filtering. The pipeline runs on Nvidia's Holoscan platform and RTX PRO 6000 Blackwell GPUs, which can pull in up to 2.4 terabits per second straight into GPU memory. The team says it detected a Crab Nebula pulsar in real time, which it calls a first. Still no alien signal, after a search running since 1960.