Field notes — week of September 7, 2026
What changed in AI-support tooling this week. Every item was checked against its source before it entered this list.
- Sierra - introduced multimodal agents (Sept 10) that shift between voice, text, and embedded visual elements - product cards, comparison tables, calendars, forms - within a single conversation, deployable across every channel without the customer repeating themselves.
- Decagon - two posts (Sept 9) cover governance for its self-improving Autopilot: Vault centralizes team-authored guidelines with version history, and separates immediate trigger-based fixes from a slower "Hillclimbing" audit track; the companion DuetBench-2 benchmark reports an 81.9% pass rate on building and revising agent procedures, and one Fortune 50 customer's resolution rate moving from 30% to 55%, with more than 85% of Autopilot's changes still live in production 30 days later.
- Intercom - pre-built data connector templates (Sept 11) for Stripe, Shopify, and Statuspage let Fin reach those tools without hand-configuring a connector. The day before, saved Inbox Views gained a CSV export that preserves columns, sort order, and a link back to each conversation.
- Anthropic - Managed Agents permission policies gained an
automode (Sept 10): the server evaluates each tool call and allows, denies, or pauses it for approval, rather than the existing all-or-nothingalways_allowandalways_asksettings.
- Lokalise - a blog post (Sept 9) describes its Proof of Value feature, which scores AI translation against a team's own approved translations rather than a generic benchmark; one cited fintech customer saw a 49% perfect-match rate with a custom AI profile versus 13% with a generic one.
- Crowdin - a case study (Sept 9) describes crypto platform EMCD scaling from 5 languages to 25 locales, including a custom JSON export built for its pluralization rules and an n8n bot that pulls Figma context before content reaches translators.