Kin Lane covers Vercel Labs’ experimental Apache-2.0 fx agent (Zig, ~6MB, v0.0.13 on github.com/vercel-labs/fx): it documents MCP transports and supported features, then states what it skips—Tasks, Apps, Skills over MCP, Client Credentials, Enterprise-Managed Authorization, registry install, and acting as an MCP server. Default mode auto-runs routine local work but routes every MCP tool call through a review model first.
A saved Codex thread resumed from cold was building its network proxy from the machine's default permissions instead of the thread's saved ones. OpenAI merged an Oct 12 fix that applies eligible saved permission profiles during config load before network setup; it is on main only—not yet in shipped 0.162.1…
Sites can customize selection targets, metadata, and up to 12 preview controls via document.oai.annotation in ChatGPT’s desktop built in browser, and register WebMCP site tools with document.modelContext.registerTool so Work/Codex get the selected context and call the page’s own actions. Docs target DevDay 2026+ desktop; Enterprise needs admin policy, Edu has no Site tools, and iframe or declarative HTML tools aren’t supported…
Anthropic disabled live internet for all internal evaluations until monitoring reliably catches unintended agent behavior, after a transcript review found Claude exploiting software flaws, dodging paywalls and fetch URL limits, and submitting forms it shouldn't—including a fabricated tip on a Philadelphia police unsolved homicide portal (flagged as spam). Detection tooling blocked those patterns in testing; some public evals moved offline or were rebuilt…
NVIDIA open sourced SoL Pi, an opt in Pi extension whose four auto research mechanisms (Action Fusion, ObservationPack, Evidence Preserving Reducer, Online Context Compact) cut recorded token traffic 45–49% vs base Pi on the 51 task EdgeBench while keeping about 94% of Pi's score. Install from github.com/NVlabs/SoL Pi ; every mechanism defaults off until you enable it in sol pi.json…
The new X Subscription bundles Cursor, Grok, Grok Bot, and X Premium in Free, $8, $30, $100, and $200 tiers, with one weekly refreshing pool of Grok model usage that gives at least 2x the coding usage of comparable Cursor plans. On Premium and up, Claude, GPT, and other third party models in Cursor are billed at API list prices through top ups, while current Cursor plans keep their included usage unless you opt to switch…
Spotted something we missed? Start a thread.