AI Agents in China: Frameworks, Vendors and Limits

Technical background

The term “agent” escalated from single-call tool use to multi-step autonomous systems during 2023-2025. The engineering consensus (see Anthropic’s “Building Effective Agents”) is that agentic patterns — tool loops, planning, multi-agent delegation — are the practical form; “agents” are not a separate model class but a system design on top of capable models.

How it works

An agent loop is: model plans a step, calls tools (search, code execution, browser, APIs), receives observations, and continues until a stop condition. Supporting machinery includes memory (short- and long-term), planning strategies, and guardrails. Multi-agent systems split roles — planner, coder, reviewer — each with its own model and tools.

Why it matters

Agents turn language models from answerers into workers: they can operate software, browse the web, and complete multi-hour tasks. Agent capability is now a first-class model attribute — several models in the China AI Hub database list it explicitly.

Chinese adoption

China’s agent ecosystem is deep. The China AI Hub agent database tracks ten Chinese agents (last verified 2026-09-22) spanning coding agents (Qwen Code, Kimi Code, MiniMax Code), frameworks (DeepSeek Harness, Qwen-Agent), autonomous agents (AutoGLM, Doubao App), and coding plans (GLM Coding Plan, Qoder). Seven of ten list MCP support; seven list computer use. On the model side, Kimi K3, Qwen3.8-Max, Doubao Seed 2.1 Pro and MiniMax M3 all list agent capability.

Pricing evidence shows how agents have been productized. GLM Coding Plan runs on GLM-5.3 / GLM-5.3-Flash with 1M-token context and bills ¥118–¥1,078/month in China (from $18/month internationally). Kimi Code runs on Kimi K3 (2.8T-parameter MoE) under Kimi membership. MiniMax Code runs on MiniMax-M3 (~428B/~23B-active MoE). Qoder’s Auto tier smart-routes across Qwen3.8-Max, DeepSeek-V4-Pro, GLM-5.3, Kimi K3 and MiniMax M3 with 1.6x/1.1x/0.3x credit multipliers.

Major Chinese companies and models

• DeepSeek — DeepSeek Harness: open-source agent framework (Cordis-based), MIT, self-hosted.
• Alibaba Cloud — Qwen Code and Qwen-Agent (open), plus Qoder.
• Zhipu AI — AutoGLM (open) and GLM Coding Plan.
• Moonshot AI — Kimi Code (open, MIT) with browser/computer use.
• MiniMax — MiniMax Code (open) and MiniMax Agent (cloud platform).
• ByteDance — Doubao App with computer-use “Work” mode.

Practical applications

Software engineering (multi-file edits, test loops), web research and form-filling, data extraction pipelines, autonomous computer operation, and customer-support workflows.

Limitations

Autonomy implies failure modes: loops that burn tokens, tool misuse, and compounding errors over long tasks. Long-horizon reliability is still the open problem. Capability labels in our database are vendor-reported unless independently benchmarked.

Deployment considerations

Use simpler workflow patterns where they suffice; add budgets, checkpoints and human approval for irreversible actions; sandbox code and browser tools; monitor token spend per completed task, not per call.

What the available evidence actually shows

The China AI Hub database establishes that agents are a broad, multi-vendor product category in China — ten tracked agents across six companies, with open-source frameworks (DeepSeek Harness MIT, Qwen Code Apache-2.0, Qwen-Agent Apache-2.0, AutoGLM Apache-2.0 code) alongside closed products (Doubao App, GLM Coding Plan, Qoder, MiniMax Agent). What the database does not yet show is independent evaluation of agent reliability: capability flags (agent, tool calling, computer use, MCP) are vendor-reported, and long-horizon reliability is untested against a common benchmark. China AI Hub analysis indicates the structural story — models gaining first-class agent capability while agent products consolidate around MCP and computer use — is well evidenced; the quality story (how reliably these agents complete multi-hour tasks) is not yet measurable from recorded data.

Future development

Agents are absorbing computer use and browser control as standard tools, standardizing inter-agent communication (MCP, A2A), and moving toward persistent memory and multi-day task execution.

Labels used above: Official fact (from Anthropic’s agent engineering guidance and the China AI Hub entity database), Vendor-reported claim (capability and pricing figures published by each vendor), and China AI Hub analysis (our synthesis, always introduced as such).