What changed in AI-support tooling this week. Every item was checked against its source before it entered this list.
-
Sierra - open-sourcedhyper-τ-bench (Sept 8), an evaluation that measures whether a model canbuild a customer-service agent rather than act as one: recover the requirements from business records and a simulated client, design it, ship it, then face held-out production traffic. Claude Opus 5 working alone cleared 23.9% of the held-out tasks; the same model paired with an engineer cleared 82.2%. The failure modes are the interesting part - developer agents opened fewer than 80 of roughly 1,700 available files, asked the client at most four questions where 20 to 25 were needed, and made cheating attempts in 17 to 42% of runs.
-
Intercom -Fin can now query a Snowflake warehouse directly (Sept 8) through an MCP connector to a server inside your own Snowflake account. It runs under a role you designate, and Intercom states that existing permissions, masking, and row-level access policies apply automatically with nothing copied or synced out.
-
Intercom - aticket assignments dataset (Sept 8) records how long a ticket sits with each team and teammate and counts the transitions between them, with a Sankey chart of the full path and a prebuilt lifecycle report. Two days later,data connector health became a chart over time rather than a snapshot, showing HTTP status codes beside latency so you can see when errors started and whether a fix held.
-
Lilt - theSeptember release (Sept 9) adds In-Context Preview, which renders a translation inside the source file's layout so overflow and truncation surface before review rather than after, and a Context Database that puts translation memories, termbases, style guides, and project metadata behind one natural-language search. An Adobe Commerce connector is now on Adobe Marketplace, and the AEM connector can deliver into Launch spaces for review before anything goes live.