Chinese researchers ship agent that finds tools mid-thought, beating OpenAI's rigid workflows

Chinese researchers from Renmin University and Xiaohongshu released DeepAgent, an AI system that discovers and uses tools dynamically within its reasoning stream. The approach abandons predetermined workflows entirely. Instead of cycling through "think, act, observe" loops like current agents, DeepAgent treats tool discovery as part of thinking itself, achieving 89% success rates on complex tasks where the best competing models hit 55%.

What's actually new

Previous agents followed scripts: plan the task, select tools, execute actions, observe results, repeat. Even OpenAI's latest research agents work with a fixed toolkit of search, code, and browsing. DeepAgent discovers tools as it reasons. Traditional agents are GPS navigation following predetermined routes. DeepAgent navigates like a local who spots shortcuts mid-journey.

The system can tap 16,000+ real-world APIs without pre-selection. When stuck, it compresses its entire interaction history into structured memory and starts fresh, preventing the death spirals that trap conventional agents in wrong exploration paths.

The architectural revolution

From Beijing's perspective: finally, agent architecture that matches how reasoning models actually think. Western labs built scaffolding around reasoning models, forcing them into rigid cycles. Chinese researchers let reasoning flow naturally.

OpenAI constrains its agents to research tools. Search, browse, code. That's it. DeepAgent pulls from movie databases, music players, robotic environments, shopping APIs, whatever the task demands. The performance gap shows: on Spotify integration tasks, DeepAgent hits 75% success versus 52% for the strongest baseline.

Enterprises get agents that actually work. Previous systems required pre-configuring tool access, defining workflows, managing state transitions. DeepAgent discovers what it needs. A film festival query triggers Vimeo search, identifies cinema professionals, constructs YouTube links, all within one coherent thought stream.

Memory as infrastructure

The memory folding mechanism solves a fundamental constraint. Long interactions explode context windows. Agents accumulate interaction history until they can't process new information. Most systems just truncate old data. Information vanishes.

DeepAgent compresses history into three memories: episodic (long-term progress), working (current state), tool (usage patterns). JSON schemas, not prose. Structure preserves precision. When an exploration path fails, the agent folds its memory and pivots. It's controlled forgetting that enables progress.

The training innovation matters. Real APIs break, rate-limit, cost money. DeepAgent simulates them with auxiliary LLMs during training. This trades perfect fidelity for stable, fast, cheap iteration. ToolPO, their reinforcement learning method, assigns credit to specific tool-call tokens rather than whole sequences. Precise feedback accelerates learning.

The scalability test

Three signals determine whether this architecture wins. First, whether memory folding scales beyond 50-action sequences to hundred-step workflows. Current tests stop at 50. Real enterprise processes run longer.

Second, API simulation fidelity. Training on LLM-simulated APIs works until it doesn't. Edge cases, authentication flows, complex state dependencies, these emerge in production. The 89% success rate assumes well-behaved APIs.

Third, tool discovery latency. Searching 16,000 tools mid-thought adds overhead. The paper doesn't report timing. Fast enough for research, but production systems measure milliseconds.

DeepAgent reveals the real bottleneck wasn't reasoning capability but architectural rigidity. Reasoning models can handle complex tool orchestration when freed from predetermined workflows. The constraint was the scaffold, not the model.

Why this matters:

• Agent architecture shifts from scripted workflows to fluid reasoning. Performance gaps prove flexibility beats structure.
• Chinese labs now lead autonomous tool use, integrating thousands of APIs while Western approaches stay limited.