Baidu Shifts Focus: From Large Models to AI Agents

🔑 Enhanced Key Takeaways

•Baidu's CEO Robin Li introduced DAA (Daily Active Agents) as a new metric to measure the output value of AI agents, predicting global DAA could eventually surpass 10 billion, shifting focus from token consumption as a cost input.
•At Create 2026, Baidu launched a suite of new AI agent products, including the general-purpose agent DuMate, the coding agent Miaoda (with app and enterprise editions), the digital human platform Baidu Yijing, and the self-evolving agent Famou Agent 2.0.
•The upgraded Baidu AI Cloud, now a full-stack AI cloud purpose-built for large-scale agent applications, features a "Token Factory" for 25% faster agent-first inference and "Harness Engineering" for advanced context management, achieving 95% task success rates in browser/Office scenarios with 23% fewer token consumption compared to OpenAI's offerings.
•Miaoda, the coding agent, can generate 90% of its own code for its app version, has an international version called MeDo (medo.dev), and has already facilitated the creation of over 1 million applications for more than 10 million users.
•Baidu's AI-powered business, encompassing AI cloud infrastructure, applications, and AI-native marketing services, reached 52% of its General Business revenue in Q1 2026, marking a significant shift in its revenue composition.

🛠️ Technical Deep Dive

Baidu's full-stack AI Cloud is specifically designed for large-scale AI agent applications, integrating AI Infra and Agent Infra.
The "Agent Infra" side features a "Token Factory" (an upgrade from MaaS Model Service) that minimizes token recalculation, leading to approximately 25% faster inference generation compared to market benchmarks. It supports major domestic models including Ernie, DeepSeek, GLM, and MiniMax.
"Harness Engineering" is a key component covering long-context management, persistent memory, tool calling, sub-agent scheduling, and Runtime capabilities, enabling 95% task success rates in browser and Office scenarios with 23% less token consumption than OpenAI's offerings.
The "AI Infra" side utilizes a layered pooling architecture for GPU memory, DRAM, and SSD, achieving KV Cache hit rates exceeding 90% and delivering long-chain agent inference performance 3x better than mainstream open-source engines.
A unified multimodal training framework provides 2x training efficiency compared to community standards.
Hardware infrastructure includes Baidu's proprietary Kunlun P800 AI chips, deployed in multiple 10,000-GPU clusters with a 97% effective training rate and over 85% linear scaling.
The Kunlun-based Tianchi 256-card supernode, launching in June 2026, is expected to offer 25% higher throughput and 50% inference efficiency improvement for various models.
Baidu's AI agents are built with four main blocks: planning, memory, tool, and action, leveraging Retrieval Augmented Generation (RAG), prompt engineering techniques like Chain of Thought (COT), Reasoning and Acting (ReACT), and Autoplan, all controlled by tracing, observability, guardrails, and performance evaluation functions.
The latest foundation model, ERNIE 5.1, powers these agents and ranks #1 among Chinese models on LMArena's text and search leaderboards.

🔮 Future ImplicationsAI analysis grounded in cited sources

Baidu's DAA metric will gain significant traction as an industry standard for measuring AI agent success.

Robin Li is actively promoting DAA as the agent era's equivalent of Daily Active Users (DAU), shifting the industry's focus from token consumption (cost) to agents' ability to deliver tangible results (value output), which addresses a persistent pain point in AI adoption.

Baidu will significantly increase its market share in the enterprise AI software and cloud services market in China.

Its full-stack AI agent capabilities, focus on native interoperability, and tools like MiaoDa directly address enterprise pain points such as data silos, workflow fragmentation, and the need for end-to-end automation, accelerating large-scale AI deployment.

The strategic shift to an 'Agent Platform' provider will lead to stronger monetization and margin recovery for Baidu.

The DAA metric is inherently tied to value creation, and Baidu's AI-powered business already constitutes the majority of its general business revenue, indicating a successful pivot towards more profitable AI-centric services.

⏳ Timeline

2013

Baidu established its Institute of Deep Learning (IDL), signaling an early commitment to AI research and commercialization.

2018

Baidu unveiled its in-house AI chip, Kunlun, demonstrating its ambition for full-stack AI architecture.

2019

The company rolled out Wenxin 1.0, its first large language model, initiating its ERNIE series development.

2023-08

Ernie Bot, Baidu's AI chatbot built on the ERNIE series, was released to the general public after regulatory approval.

2024-11

Baidu unveiled Miaoda, a no-code AI-powered programming tool, at Baidu World 2024, emphasizing AI agents as the mainstream form of AI applications.

2026-05

At Baidu Create 2026, the company introduced DAA as a new core metric, launched a new generation of AI agent products, and announced a comprehensive upgrade of its AI Cloud to support large-scale agent deployment.