Field notes — week of August 31, 2026
What changed in AI-support tooling this week. Every item was checked against its source before it entered this list.
- Anthropic - the
ant applycommand (CLI 1.30.0, Sept 3) lets teams declare agents, environments, skills, memory stores, and deployments as files in a repository and reconcile them with a reviewable plan, writing aclaude-lock.jsonlockfile so repeat runs update the same resources instead of creating new ones.
- Salesforce - the Winter '27 editions (Sept 3) restructure the product line into three tiers - Core, Advanced, and Max - that bundle AI, Slack, analytics, security, and support together, making Agentforce native to every tier instead of a separate add-on.
- Crowdin - its August recap (Sept 4) covers a CLI rewritten in TypeScript for faster startup, four new automation Skills including one for i18n setup, an AI Pipeline builder in Copilot, and a new Adobe Experience Manager connector.
- Lokalise - a blog post on AI translation quality (Sept 1) argues human review alone can't scale, and lays out a three-layer framework - pre-production assessment, in-production scoring, and post-edit analytics - for measuring it instead.
- Decagon - two engineering posts (Sept 3): one describes disaggregating prompt processing from token generation to cut GPU-hours by roughly 80% for the same traffic; the other describes an audio-native system for detecting mid-call speaker changes that combines speaker-embedding models with a multimodal LLM.