DeepSeek Harness: A New Agent Architecture Paradigm

DeepSeek Harness introduces Cordis composability model, inspired by Koishi's plugin lifecycle, enabling dynamic addition/removal of capabilities without breaking the system. This is a paradigm shift in agent architecture, analogous to Unix's 'everything is a file' abstraction. DeepSeek is having its second DeepSeek moment, moving beyond coding assistant to a formalized agent framework. This could redefine how reliable agents are built. Security assessment reveals up to 25.5% attack success on indirect prompt injection, challenging safety assumptions; Agent Lightning v1.0 achieves 14.6-point gain on SWE-bench with harnessed agentic RL; HarnessRisk provides lifecycle safety benchmark; Harness the Memory evaluates memory substrates; fx tiny open coding agent harness released. DeepSeek open-sourced Harness under MIT license, model-agnostic plugin architecture, V4-Pro-0813 benchmarks (needs independent verification). New related papers: Zetta ζ (closed-loop embodied harness for self-evolving physical intelligence), SemaPLC (verification-gated agent harness for PLC code generation), Looped Language Models (improve compositional tool calling), SPADE (self-play in adaptive synthetic executable environments for open-ended skill acquisition). New benchmarks: LHTB (long-horizon terminal agents), FM-Bench (long-horizon management with competing agents). New open-source harness: Grok Bot (works with any harness, real browser, agent-human handoffs). These developments reinforce the harness paradigm as a key direction for reliable agents. New: SkillGate paper addresses selector credit starvation in skill-based agents, achieving 12.4% lift on 9B policy, directly relevant to agent skill pool scaling issues. New: A study shows agents lock into a training strategy at step one and cannot reconsider, challenging autonomous self-improvement — relevant to harness limitations.