Training agentic policies with real-world tools

Training with real-world web-search, coding, and notebook tools for tasks such as search and code engineering.

Also called real-world tools.

source: 1
model: 1
lab adopt it: 1
strongest: used

How sources treat it

One count per evidence span, weakest treatment to strongest.

used 1

Documented in

DeepSeek-V3.2

Evidence

1 span quoted from the sources, strongest treatment first.

For tasks such as search, code engineering, and code interpretation, we employ real-world tools, including actual web search APIs, coding tools, and Jupyter Notebooks.

useddata curationin DeepSeek-V3.2DeepSeek

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models·DeepSeek·2025-12-02

Filed alongside

Other methods under post-training :: agentic post-training.

Agentic reinforcement learning Multi-harness training Agentic tool-use training Environment hardening Reinforcement learning on synthetic agentic data Repair-agent environment correction loop Agentic post-training Autonomous Execution Tasks (AET)Concurrent multi-environment post-training Container-level network isolation Environment preparation to prevent solution leakage Environment scaling for realistic training tasks Generate-verify-refine loop Harness-optimized training Mock applications for personal-assistant reinforcement learning Multi-agent task decomposition and coordination Multi-environment reinforcement learning Multi-environment reinforcement learning from verifiable rewards Multi-harness trajectory training with native behavior preservation Multi-turn collaboration simulator Production-harness-matched reinforcement learning Python tool use in chain-of-thought Re-post-training for agentic capabilities Reinforcement learning in an agent harness