LangChain adds ScaleDown constrained decoding models for agent context

Meta says Muse Wonder models are too large for many people to use and carry potential safety concerns, so they will not be open-sourced. Open weights for Muse Spark are coming soon.

Anthropic migrated its monorepo off git to an in-house version control system called Scion in five months, after daily merges grew from ~2,500 in May 2026 to over 30,000 by September. It also shipped a code review surface called Antique, code search, and a source control IAM system in the same period.

MCP support now ships in the main langchain.mcp package, built on FastMCP for the 2026-07-28 spec, replacing the separate langchain-mcp-adapters install. Elicitation is surfaced as a LangGraph interrupt and tool lists are now cacheable, matching the spec's stateless core.

LangChain's Deep Agents evals found GLM-5 (z.ai) and MiniMax M2.7 score similarly to closed frontier models on file operations, tool use, and instruction following. Outputting 10M tokens/day costs ~$250/day on Opus 4.6 versus ~$12/day on MiniMax M2.7 — about $87k annually.

LangChain automated migrating inline docs code samples into standalone, CI-tested files using its Deep Agents CLI, so broken examples fail the build. The manual path otherwise requires extracting snippets, adding setup/teardown, snippet markup, and CI wiring.

Jensen Huang said at a Paris event that Nvidia will announce $2 billion in European investments over the next week, on a tour that also takes him to London and Berlin.