Baidu's DAA Metric, Amazon's Agentic Shopping Pivot, and Self-Hosted Coding Agents Go Beta
News
Baidu Proposes "Daily Active Agents" as the New Industry Metric at Create 2026
Baidu CEO Robin Li made a pointed argument at today's Create 2026 conference: tokens measure cost, not value, and the right scoreboard for the agent era is Daily Active Agents (DAA), the equivalent of mobile's DAU. Alongside the metric push, Baidu shipped DuMate (general-purpose agent with mobile app), Miaoda (a coding agent that generates 90% of its own code), Famou Agent 2.0 (topped MLE-Bench, winning 9 of 15 hardest tasks), and Yijing (a 12-language digital human platform). Watch whether DAA catches on as a framing device, because whoever defines the measurement tends to win the narrative.
Amazon Kills Rufus and Launches Alexa for Shopping as a Full Agentic Commerce Layer
Amazon retired its Rufus chatbot and folded it into Alexa for Shopping, which now lives directly inside the Amazon search bar and can compare products, track prices across a full year of history, and use "Buy for Me" to complete purchases on other retailers' sites. This is less a rebrand and more a structural bet: Amazon is trying to ensure that the agent mediating the next billion shopping decisions runs on Amazon infrastructure rather than ChatGPT, Perplexity's Comet, or Google's Gemini. If agentic commerce is where retail attention flows, Amazon just declared it's not giving that surface to anyone else.
Tools
Coder Ships Self-Hosted AI Coding Agents Beta: No Code Leaves Your Network
Coder launched Coder Agents in beta, a model-agnostic coding agent that runs entirely on the customer's own infrastructure so no source code, prompts, or model traffic touches a third-party cloud. With 70% of companies currently deploying agents on infrastructure not designed for them (per Coder's own research), this directly addresses the compliance and data residency wall that has blocked regulated industries from adopting AI-driven dev workflows. Full feature access runs free through September, which is a smart land-and-expand move for teams that need to pilot before committing.
Analysis
Baidu Create 2026: The AI Industry Shifts from Model Competition to Agent Competition
Robin Li's Create 2026 keynote framed a clear inflection: the products going viral in 2026 are not better models, they are better agents built on top of models. "For the first time, what really made AI go viral was not the model, but the application," he said, signaling that differentiation is now at the application and orchestration layer rather than pretraining. For builders, this validates the thesis that infrastructure, memory, task execution reliability, and agent UX are the new moats, not raw parameter count.
The Next Web's breakdown of the Alexa for Shopping launch zeroes in on the structural change: the AI now sits inside the default search flow rather than behind a separate icon, reframing the Amazon search bar itself as a conversational agent interface. This piece also surfaces the competitive context cleanly, including Amazon's ongoing legal battle with Perplexity over its Comet shopping agent and the implication that Amazon's $56B ad business depends on staying the first surface buyers touch. Worth reading alongside the CNBC announcement for the full strategic picture.
Research
This paper from ByteDance Seed and Renmin University introduces Agent-World, a framework that autonomously mines real-world databases and MCP tool ecosystems to synthesize over 19,000 validated tools across 2,000 environments, then uses multi-environment RL with a self-evolving arena to continuously train agents against their own capability gaps. The 8B and 14B models trained on this corpus beat strong proprietary baselines across 23 benchmarks, with Agent-World-14B matching DeepSeek V3.2-685B on BFCL-V4 at a fraction of the parameter count. The key insight is that scalable, diverse training environments may be as important as model scale for building general agent capability.
Stay Ahead
Delivered each morning.