Field notes — week of September 14, 2026

Anthropic's Messages API added an on-demand conversation compaction feature in beta on September 14, 2026, using the compact-2026-09-04 header to return a signed block that summarizes prior messages so they can be resent in place of the originals while recent turns stay word for word. The week's other AI-support and localization updates included Sierra achieving AIUC-1 certification on September 17 after an independent Schellman audit and AI Underwriting Company testing, DeepL Voice reaching general availability on September 15 with a 4% translation error rate in cited testing versus a 17% average for competing platforms, and Crowdin reporting Sinch scaled to 5 million words translated in Q1 2026 with AI automating 65% of translations via Google Gemini 3.1 Pro.

Field notes — week of September 14, 2026 What changed in AI-support and AI-localization tooling this week. Every item was checked against its source before it entered this list. - Anthropic - the Messages API can now compact a conversation on demand https://platform.claude.com/docs/en/build-with-claude/compaction compact-on-demand-with-the-compaction-parameter Sept 14 , in beta with the compact-2026-09-04 header; sending the top-level compaction parameter returns a signed block summarizing prior messages, which can be resent in place of them on later requests while recent turns stay word for word. - Sierra - achieved AIUC-1 certification https://sierra.ai/blog/sierra-achieves-aiuc-1-certification Sept 17 after an independent audit by Schellman and testing by the AI Underwriting Company, which probed its chat and voice agents for manipulation attempts and unauthorized data access; the evaluations recur quarterly with a full audit each year, on top of Sierra's existing SOC 2 Type II, ISO 27001, and ISO 42001 certifications. - Decagon - two posts this week: GEPA-GAN https://decagon.ai/blog/gepa-gan-teaching-ai-to-sound-human Sept 15 describes an adversarial prompt-optimization method - a generator, a judge, and a reflector role, with no model weights changed - that raised a "human fool rate" from 30% to 34% for the simulated customer messages used to test agents; Generative Analytics https://decagon.ai/blog/generative-analytics Sept 18 lets teams generate charts from conversation data by asking in natural language, plus trigger-based "Action Trackers" - one retailer used the feature to find its agent couldn't find a suitable product for 26% of customers. - DeepL - DeepL Voice https://www.deepl.com/en/blog/deepl-voice-preserves-your-voice-across-languages Sept 15 moved from beta to general availability and now preserves each speaker's tone and vocal character across languages in real time, in a unified desktop app for Zoom, Teams, and Google Meet covering 12 languages; independent testing cited by DeepL found a 4% translation error rate versus a 17% average for competing platforms. - Lokalise - introduced Vantage https://lokalise.com/blog/introducing-vantage/ Sept 16 , a no-code workspace for localizing marketing and long-form content with native integrations for Contentful, Webflow, Braze, Strapi, Google Docs, and SendGrid; Lokalise reports customers cutting new-market launch time from four weeks to under two days. - Crowdin - a case study https://crowdin.com/blog/sinch-localization-with-crowdin-and-undertow Sept 17 describes Sinch scaling to 5 million words translated in Q1 2026 alone - more than its full 2025 volume - with AI now automating 65% of translations via Google Gemini 3.1 Pro through Crowdin's bring-your-own-key feature, and 44% of AI output still receiving human proofreading. - WPML - WPML 5.0 https://wpml.org/announcements/2026/09/wpml-5-0/ Sept 16 consolidates settings into one admin menu, runs translation in the background without an open browser tab, and moves its PTC AI engine to pay-per-page billing where edits to an already-paid page are free; taxonomies gain translation support and languages can now be split by country pair.