Company@X

Signal of the Day

Google’s rapid launch of Gemini 3.7 Flash just three weeks after its predecessor—offered at a 50% introductory price cut to run agentic developer pipelines—signals a dramatic acceleration in model release cadences and aggressive price competition for production-ready agent workloads.

Key Announcements

Google · Source Google has launched Gemini 3.7 Flash, bringing major performance upgrades to coding, software engineering, and multi-step agentic workflows. To drive immediate developer and enterprise adoption, Google is offering an introductory price of $0.75/1M input and $3.75/1M output tokens (a 50% discount) through the end of 2026, alongside launching “Sheets canvas” to prompt static data directly into interactive mini-dashboards. Additionally, Google joined as a core maintainer of Agent Plugins 1.0.0 to standardize how skills and database tools are bundled so they can run seamlessly across compatible workspaces.

Cursor · Source Cursor has introduced “builds,” continuously prepared background development environments that reduce cloud agent startup times by 3x, dropping latency from minutes to seconds for users like Faire and Descript. To expand its core execution capabilities, the company also acquired the Firetiger team to build agents capable of following code changes directly into production and autonomously fixing live bugs. This performance optimization and talent acquisition represents a major push to make autonomous, long-horizon developer tasks highly resilient and practical.

Exa · Source AI-native search engine Exa has surpassed a massive index milestone, currently serving 80 billion pages, tracking 1.4 trillion URLs, and targeting Google-scale capacity by early 2027. Founder Will Bryk notes that the rapid rise of agentic web search is driving immense, spiky traffic—often requiring dozens of sub-searches per query—which is forcing Exa to provision for Bing-scale peak QPS ahead of schedule. Bryk forecasts that within two years, AI agents will execute searches at millions of queries per second, requiring retrieval infrastructure that eclipses existing search scales in both dimensions.

Sequoia Capital · Source At Sequoia Capital’s “Own Your Intelligence” event, partner Sonya Huang highlighted that Jevons Paradox is driving up gross margins for AI application companies as intelligence becomes cheaper to meter, while model companies see improved cohort performance. The event heavily focused on closing the “Experience Gap” where models possess raw IQ but lack specific enterprise tenure, highlighting the critical role of custom agent harnesses. LangChain founder Harrison Chase proposed that custom harnesses are essential when operating out of distribution from a model’s base training data.

Waymo · Source Waymo has partnered with the Las Vegas Raiders and Allegiant Stadium to bring autonomous rides to Las Vegas, establishing the first-ever partnership between a professional football team and an autonomous vehicle company. The deal includes a dedicated stadium pickup zone to handle autonomous rides during concert nights and game days.

Decawork · Source Decawork has launched with Y Combinator backing to provide a centralized platform for enterprise IT departments to deploy, govern, and maintain internal AI agents. Founded by AI veterans from NVIDIA, OpenAI, and Meta, Decawork addresses the security chaos of team-built agents by enforcing scoped credentials, mandatory sign-offs, and administrative approvals for sensitive actions.

Prime Agent · Source A Prime Agent setup utilizing a Recursive Language Model (RLM) and a Continual Harness has successfully enabled Opus 5 to beat the human-expert baseline on the ARC-AGI3 evaluation for the first time, scoring 95.5% versus 95.4%. This breakthrough offers strong evidence for the power of agentic harnesses over simple model wrappers, prompting Y Combinator to dedicate its next Paper Club on August 26 specifically to harness architectures.

Also Noted

• OpenCode (Source): Grok 4.6 has officially launched on OpenCode Zen and is available for SuperGrok subscribers, offering frontier-adjacent intelligence for agentic CAD work at a lower cost.
• Parasma (Source): Known for training human brain cells to play Doom, Parasma has now successfully trained them to predict words in a sentence, advocating for biological compute as an ultra-power-efficient alternative to silicon.
• LemonSlice (Source): The company has integrated its advanced interactive AI avatars with Agora’s ultra-low-latency real-time video and voice platform, enabling face-to-face AI agents for customers like Whatnot and Talkspace.
• DeepReach AI (Source): Backed by Y Combinator, DeepReach has launched a platform to help entrepreneurs build local real-world data businesses tailored to Physical AI and robotics.

🎧 This digest would make a great brief audio overview if you want a quick podcast-style breakdown of today’s agentic engineering announcements.