Baidu pivots strategy from AI models to AI agents

🔑 Enhanced Key Takeaways

•Baidu introduced a comprehensive suite of new AI agent products at Create 2026, including DuMate (a general-purpose agent with real-time mobile and PC synchronization), Miaoda (a coding agent capable of generating 90% of its own code, with an international version called MeDo), Baidu Yijing (an upgraded digital human platform supporting 12 languages for content creation), and Famou Agent 2.0 (a self-evolving agent for enterprise tasks like production scheduling and logistics planning).
•Baidu CEO Robin Li proposed 'Daily Active Agents (DAA)' as a new core metric for the AI industry, asserting that it more accurately measures value and task completion compared to traditional 'Daily Active Users (DAU)' or token consumption, and predicted global DAA could eventually exceed 10 billion.
•The strategic pivot to AI agents is supported by a 'new full-stack AI cloud' specifically designed for large-scale agent applications, featuring infrastructure upgrades across Agent Infra and AI Infra, and leveraging a dedicated cluster powered by Baidu's proprietary Kunlunxin AI chip for training models like the ERNIE 5.1 series.
•The shift in focus from foundational models to AI agents is driven by the realization that the primary value and virality of AI products now stem from their ability to autonomously 'get things done' and execute tasks, rather than solely from the underlying model's intelligence.
•Baidu's AgentBuilder platform facilitates the creation and deployment of AI agents with low-code/no-code options, allowing developers and businesses to build customized AI-powered applications that can integrate with Baidu's own ERNIE models and potentially external models, offering features like embedded links and conversion tools for e-commerce.

🛠️ Technical Deep Dive

Baidu's AI agent strategy is underpinned by a "Chip, Cloud, Model, Agent" full-stack architecture.
The chip layer utilizes the Kunlun Core P800, which has been validated for large-model training and inference.
Baidu AI Cloud has undergone comprehensive upgrades, including optimizations for long-context management, to serve as a full-stack AI cloud for large-scale agent applications.
The foundational large language models are from the ERNIE series; ERNIE 5.1, the latest iteration, achieved flagship-level intelligence by reducing total parameters by two-thirds compared to ERNIE 5.0 and activating only half during inference, leading to significant cost efficiency.
A typical AI agent architecture includes four main building blocks: planning, memory, tool, and action, utilizing Retrieval Augmented Generation (RAG) to access external information and tools during inference.
Agent performance can be enhanced through prompt engineering techniques such as Chain of Thought (COT), Reasoning and Acting (ReACT), and Autoplan.
Baidu employs a four-stage reinforcement learning system called MOPD (Multi-Teacher On-Policy Distillation) in its post-training pipeline for models like ERNIE 5.1.
The AgentBuilder platform provides zero-code and low-code modes, enabling agents to be deployed across various Baidu products, including Baidu Search, Baidu Maps, and ERNIE Bot.

🔮 Future ImplicationsAI analysis grounded in cited sources

The shift to AI agents will significantly democratize software creation, enabling non-technical users to build applications.

Baidu's Miaoda coding agent and AgentBuilder platform emphasize no-code/low-code development, allowing users to create applications and agents with natural language or minimal coding.

The 'Daily Active Agents' metric will become a critical industry benchmark, shifting focus from user engagement to AI-driven productivity and task completion.

Robin Li explicitly proposed DAA to measure the actual task completion loops of AI agents, arguing it better reflects value creation and productivity than traditional DAU or token consumption.

Enterprise organizational structures will flatten, with managers focusing more on 'alignment' as individuals are augmented by 'super individuals' (human + AI agent formations).

Robin Li's 'AI evolution theory' suggests individuals will become 'super individuals' with AI assistance, leading to organizational restructuring where managers oversee larger teams and focus on aligning human-agent collaborations.

⏳ Timeline

2013

Baidu established its Institute of Deep Learning (IDL), marking its early, foundational investment in AI.

2019

Baidu released ERNIE 1.0, its first large language model, laying the groundwork for its foundational models.

2023-08

ERNIE Bot, Baidu's advanced large language model chatbot, was released to the general public.

2024-04

Baidu introduced its AgentBuilder platform, enabling low-code/no-code AI agent development.

2026-01

Baidu released ERNIE Bot 5.0, a significant iteration of its foundational model with advanced multimodal capabilities.

2026-05

At Create 2026, Baidu officially announced its strategic pivot to AI agents, unveiled new agent products, and proposed 'Daily Active Agents (DAA)' as a new industry metric.

Baidu pivots strategy from AI models to AI agents

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (25)

👉Related Updates