
DeepSeek opens beta testing for Harness, its agent layer for autonomous multi-step code
DeepSeek is recruiting open-source developers to beta test DeepSeek Harness, software that turns its models into agents able to execute multi-step code and act autonomously, led by Cui Tianyi — ex-TSY Capital co-founder and Jane Street engineer — hired in March 2026. No release date was disclosed and the team was still described as severely short-staffed in June, so the agent layer trails the model work; the competitive pressure remains the price line, where Liang Wenfeng's cheap-capable-model bet and the V4 Flash release sent another cost shock through Silicon Valley.
Source: scmp.com ↗
The Hangzhou-based firm was looking for open-source project developers to join beta testing for DeepSeek Harness
scmp.com
Why this matters
- → DeepSeek escalates agent capabilities race while undercut pricing pressures US AI vendors
- → Shortage signals execution risk despite model prowess; harness lags V4 development
- → Cheap-capable bet now extends beyond inference to autonomous multi-step workflows
Cheap meets autonomous