Recursive released Flint, an Apache 2.0 open-weight mixture-of-experts coding model with 39.8B total parameters and roughly 2.5B active per token, built in under four months largely by the company's own automated research system and free on OpenRouter at launch.

• Runs at 125-plus tokens per second on a MacBook Pro, with weights on Hugging Face and reported scores of 89.7% on SWE-bench Pro v2 and 74.2% on SWE-bench Verified.
• Built by the Eureka Machine, Recursive's automated research system, which handled nearly every step of development, making Flint the first public product from the $650M-backed lab.

mabl introduced a new agentic testing harness that lets engineering and quality teams set standards and guardrails for their coding agents and preview the impact of any change from the terminal before a pull request is opened.

• Teams define guardrails independently of whichever coding agent produced the code, giving QA an unbiased check grounded in how the application actually behaves in production.
• Terminal-first feedback loop means developers see verification results without leaving their workflow, closing the gap between agent output and a mergeable pull request.

Zoom shipped a unified context layer that ingests meeting conversations and connects them to downstream tasks, CRM records, and calendar actions so that commitments made in calls automatically move into completed work without manual follow-up.

• Context layer spans Zoom's full stack, linking AI Companion summaries to action items, deal updates, and scheduling across integrated tools in a single pass after a meeting ends.
• Headless AI strategy built with Anthropic means the same context can surface inside third-party tools, including Salesforce workflows and Claude-based agents, rather than staying locked in Zoom.

CaptivateIQ moved its Rev Planning Agent and Comp Builder Agent to general availability, letting revenue operations teams run plan changes in days rather than weeks using the organization's own compensation logic and rules, with every output auditable.

• Rev Planning Agent models quota and capacity scenarios autonomously, cutting the cycle time for territory and headcount changes that previously required weeks of spreadsheet work.
• Comp Builder Agent reads existing plan rules to generate and modify compensation structures, keeping every calculation tied to source logic so finance and legal can audit the output.

Sierra released tau-i18n, an extension of its tau-Voice evaluation framework that measures voice agent performance across Spanish, Brazilian Portuguese, Hindi, Korean, and Mandarin using the same airline, retail, and telecom tasks as the English baseline.

• Accent-driven drop identified in prior research prompted the expansion; tau-i18n measures not just task completion but service quality across language and dialect variation.
• Same task corpus across all six languages means enterprise teams can benchmark multilingual deployments on a consistent standard rather than building per-language evaluation sets.