Baidu Qianfan Agent Development Platform

From PRIMO.ai
Jump to navigation Jump to search

YouTube ... Quora ...Google search ...Google News ...Bing News


Baidu Qianfan (百度千帆) is Baidu AI Cloud's enterprise platform for building, deploying, and running AI agents. Its official full name is the Baidu Qianfan Large Model Service and Agent Development Platform. It bundles four things in one console:

  • Model access: Baidu's ERNIE family plus more than 150 third-party and open-source models.
  • Agent development: A layer that grew out of Qianfan AppBuilder.
  • Tools and MCP: A marketplace anchored by Baidu AI Search.
  • Enterprise services: Data, operations, and governance capabilities.

Baidu launched Qianfan in March 2023 and rebranded it around agents in October 2025. Baidu reports more than 460,000 enterprise customers and more than 1.3 million agents built on the platform. Baidu also describes Qianfan as the "Agent Infra" behind its own agent products, including the DuMate general-purpose agent.


Baidu Qianfan treats the agent, not the model, as the unit of enterprise AI: one platform where companies pick from ERNIE and 150+ other models, connect them to Baidu Search and MCP tools, and ship finished agents into WeChat, Baidu Search, and their own systems.


  • The Mechanism: Baidu organizes Qianfan's Agent Infra into five layers: model services, tool services, agent development services, data services, and an agent runtime environment. The agent engine supports three orchestration styles (autonomous planning, workflows, and multi-agent collaboration). Developers can build in zero-code, low-code, or full-code modes. The development toolchain provides:
    • RAG and knowledge bases, including multimodal RAG for images and tables
    • a visual workflow canvas
    • a UI Builder
    • prebuilt components such as Baidu AI Search, document understanding, image understanding, and speech recognition
    • full-link monitoring that shows the inputs, outputs, and performance of every step in a run

Finished agents can be published to Baidu's Miaoda app builder, a web page, Baidu Search, WeCom, WeChat Official Accounts, WeChat customer service, or an embedded website widget.

  • The Physics: Enterprise agents run under hard limits on cost, latency, and time. Qianfan's engineering targets each one (the figures below are Baidu-reported):
    • Cost: Chain-of-thought control matches short reasoning paths to simple questions and saves multi-step reasoning for hard ones. A distributed KV cache raised average throughput by 61% on a 10,000-card inference cluster. An "Active Cache" mode cuts inference costs for long, multi-turn agent sessions by up to 80%.
    • Latency: The upgraded agent engine cut average task latency by 20% and long-task time by 40%.
    • Time budgets: Workflow conversation timeouts rose from 6 to 30 minutes in October 2025. The Deep Research agent offers a "performance-first" mode (at most 5 search rounds, about 10 minutes) and a "quality-first" mode (about 10 rounds, 20 to 30 minutes).
    • Session lifetimes: A conversation ID and its uploaded files expire after 7 days.
  • Why This Matters: Qianfan is Baidu's main commercial bet on enterprise agents.
    • Exclusive search tool: Its signature tool is Baidu AI Search, which brings China's largest search index into agents as a real-time grounding source. Baidu reports tool calls in the tens of millions per day.
    • Open model shelf: The platform is deliberately model-neutral, hosting DeepSeek, Kimi, GLM, MiniMax, and Qwen models alongside ERNIE.
    • Coding-agent compatibility: Its Token Plan subscriptions connect to coding agents such as Claude Code, Codex, Cursor, OpenCode, and DeepSeek Harness.
    • Business traction: In Q2 2026, Baidu said revenue from external customers' token usage on Qianfan grew more than ninefold year over year.
    • Strategy: At Baidu Create 2026, CEO Robin Li proposed Daily Active Agents (DAA) as the industry's key success metric, positioning platforms like Qianfan as the factories that produce those agents.

Agent Types

Qianfan's Agent development studio offers four application types:

Agent Type How It Works Best Used For
Autonomous Planning Agent (自主规划Agent) The model plans its own steps using role instructions, a chosen model, knowledge bases, components, memory, and databases. It can be published one-click as a Miaoda plugin. Open-ended assistants and tasks where the steps aren't known in advance.
Workflow Agent (工作流Agent) A visual canvas of nodes, including LLM, agent, component, API, MCP Server, CFC function compute, intent recognition, branching, code, loops, knowledge base, database, and memory variables. Deterministic business processes such as order handling, customer service flows, and insurance renewals.
Multi-Agent Collaboration Agent (多智能体协同Agent) An orchestrator distributes work across several specialized agents and is compatible with the A2A protocol. Qianfan 4.0 marketed it as a way to build a custom "Manus." Complex, multi-step tasks such as enterprise deep research.
Interactive Writing Agent (交互式写作Agent) Template-driven long-form generation, with smart editing and document export. Reports, official documents, and other structured writing.

Official ready-made agents include:

  • Deep Research, which produces cited, structured reports from web and uploaded sources and can export the full research "working papers" in Excel format
  • an AI Assistant for general and image-based Q&A
  • a Browser Use agent based on the open-source BrowserUse project
  • a Code Interpreter agent that runs Python in a sandbox
  • a configurable Agent Engine that routes requests to curated MCP servers

Getting Started

  • Sign in: Open the Qianfan console through Baidu AI Cloud at https://console.bce.baidu.com/qianfan/, then create an API key.
  • Call models: Qianfan offers OpenAI-compatible REST APIs and multi-language SDKs, plus a Responses API (with function calling and built-in knowledge-base search) and Anthropic-format API support.
  • Build agents: Use the Agent development studio in zero-code or low-code mode, or go full-code with the Agent development APIs:
    • create a conversation
    • upload session files (one at a time)
    • send messages to an agent by its app_id
    • manage knowledge bases and document chunks
  • SDK: The AppBuilder SDK (baidubce/app-builder on GitHub) exposes Qianfan models, 40+ Baidu ecosystem components, retrievers built on Baidu Vector Database and Baidu ElasticSearch, and deployment helpers.
  • Ecosystem integrations: LangChain and LlamaIndex both ship Qianfan connectors.
  • Tools and MCP: Browse the Tools and MCP Square for official and third-party MCP servers. Developers can also publish their own MCP servers on Qianfan, which Baidu Search can index for discovery.
  • Coding plans: Token Plan (personal and enterprise editions) replaced the earlier Coding Plan in July 2026. The enterprise edition is a seat-based monthly subscription metered in points, with team management and usage analytics.

Known Limitations

  • API gaps: Workflow components that contain memory nodes or the Baidu Search component cannot be called through the API or SDK.
  • Expiring sessions: Conversation IDs and session files expire after 7 days, so long-lived integrations must regenerate them.
  • Model churn: Third-party models rotate on and off the platform. For example, Kimi-K2.5 and MiniMax-M2.5 were retired from the V2 inference API on July 9, 2026.
  • Shifting plans: Billing products change quickly. Coding Plan renewals stopped on June 25, 2026, in favor of Token Plan.
  • Language: Most detailed documentation is in Chinese. An English product overview exists on Baidu AI Cloud's international site.
  • Vendor-reported metrics: Agent counts, customer totals, latency gains, and cost reductions come from Baidu's own announcements.

Recent Developments

  • March 2023: Baidu AI Cloud launched Qianfan.
  • December 20, 2023: Qianfan AppBuilder, the AI-native app development workbench, opened to all users.
  • August 28, 2025: At the Baidu Cloud Intelligence Conference, Baidu released Qianfan 4.0, which added:
    • an expanded library of 150+ models
    • an RFT (reinforcement fine-tuning) toolchain
    • multimodal RAG
    • an exclusive Baidu AI Search MCP Server
    • a new multi-agent collaboration mode
  • October 16, 2025: Baidu rebranded the platform as "Baidu Qianfan," refocused on enterprise production use. The upgrade added chain-of-thought control and the Active Cache mode. The reported scale at the time was 460,000+ enterprise customers and 1.3 million+ agents.
  • November 2025: Qianfan added:
    • ERNIE-5.0-Thinking-Preview, Baidu's natively omni-modal model
    • an official Agent Engine in the app square
    • the Responses API, with function calling
  • December 11 to 12, 2025: At an agent conference in Chengdu co-hosted with NVIDIA, Qianfan announced Active Cache pricing, an MCP Square, and new official agents for Deep Research, Browser Use, and Code Interpreter.
  • February 4 to 5, 2026: Qianfan's Deep Research agent (Qianfan-DeepResearch Pro) took first place on the DeepResearch Bench leaderboard.
    • Baidu commercialized Deep Research at 2.5 yuan per call, with 50 free calls per account.
    • Baidu launched a Qianfan developer ambassador program.
    • A hybrid-cloud version of Deep Research was announced as coming soon.
  • February 13, 2026: Autonomous planning agents gained one-click publishing to Miaoda, and Qianfan added a quick-setup page for OpenClaw.
  • March 20, 2026: Baidu released the DuMate MVP, a desktop agent for macOS and Windows.
    • It shipped with a preinstalled security sandbox and second confirmation for high-risk actions.
    • It supported interrupting and resuming conversations and running multiple tasks in parallel.
  • May 2026: Baidu released ERNIE 5.1, which Baidu says ranked first among Chinese models on LMArena's text and search leaderboards. At Baidu Create 2026 (May 14):
    • Robin Li proposed the DAA metric.
    • Baidu launched DuMate's mobile edition and Famou Agent 2.0.
    • Baidu repositioned Baidu AI Cloud as a full-stack cloud built for large-scale agent applications.
  • June 2026: Baidu's Qianfan team published the DuMate-DeepResearch technical report. It describes a multi-agent research system built on the "Qianfan Agent Foundry," using graph-based planning, recursive searcher agents, and rubric-guided reasoning.
  • June 25 to July 2026: Coding Plan renewals ended, and Token Plan personal and enterprise editions took over. Supported tools include Claude Code, Codex, Cursor, OpenCode, OpenClaw, Qwen Code, Cline, Kilo, and DeepSeek Harness.
  • August 2026: Several changes landed within weeks:
    • Earnings (August 18): Baidu reported Q2 2026 results, with AI Cloud Infra revenue up 50% and Qianfan external token-usage revenue up more than ninefold year over year.
    • Benchmark (August 28): Baidu launched DuMateBench, an agent evaluation with 200+ office tasks across six categories.
    • Reorganization: TechNode reported that Baidu Cloud renamed its platform product division as an intelligent-agent division and moved the MaaS and Qianfan businesses into a separate infrastructure division.

Load video
YouTube
Load video
YouTube

Retrieved from "https://primo.ai/w/index.php?title=Baidu_Qianfan_Agent_Development_Platform&oldid=45917"