Which AI writes the better take? You decide — blind.

Manufact is turning its mcp-use SDK (10k GitHub stars, 7M downloads) into a hosted platform for deploying, monitoring, and debugging MCP servers, with cross-client testing across ChatGPT and Claude plus audits against ChatGPT Apps Store and Cloud Connectors publishing requirements. If you're shipping MCP tools, this offers a managed path to production—trace/replay of live MCP traffic and pre-submission compliance checks—instead of hand-rolling hosting and per-client validation yourself. The tradeoff is betting your deployment layer on a young YC startup's cloud rather than owning the infra.

Simon Willison's LLM library has quietly become an agent framework, and a working Claude-Code-style coding agent (file read/edit, shell exec with process-tree-kill on timeout, glob/regex search, gitignore-aware listing) was built on it in two prompts and shipped to PyPI—runnable today via uvx with a Python API like CodingAgent(model=..., root=..., approve=True).run(...). Practically, this means you can embed a scriptable, permission-gated coding agent (with --allow allowlists and --yolo modes) directly into Python workflows without adopting a heavyweight vendor SDK, though it's a slop-alpha so treat it as a prototype, not production tooling.

Datasette Agent’s SQL prompt was causing models to guess nonexistent column names because the schema context listed only table names while discouraging extra describe_table calls. For production text-to-SQL agents, include column names in the initial schema context or relax tool-use constraints, otherwise you’ll pay for avoidable failed queries, retries, and lower answer reliability.

Anthropic is exploring a Samsung-built custom chip, following OpenAI's Broadcom "Jalapeño" inference processor, but this is early-stage with no defined use case, power target, or server integration yet—Anthropic explicitly reaffirms its diversified stack across Google, Amazon, and Nvidia. Practically, nothing changes your capacity planning today: keep provisioning around existing TPU/Trainium/Nvidia availability, and treat this as a multi-year signal that frontier labs are hedging against Nvidia lock-in rather than an imminent hardware shift you can build against.

Every major model vendor now runs a forward-deployed engineering arm—Microsoft ($2.5B, 6,000 engineers), AWS ($1B), plus OpenAI and Anthropic—signaling that enterprise AI value has shifted from the model to the integration labor around it. If you're shipping AI internally, expect these teams to compete with (or displace) your systems integrators and consultants, with vendors increasingly owning the deployment layer and the outcome accountability that comes with it—which also means more lock-in to a single vendor's tooling.