DeepSeek & Peking University Paper Unveils Self-Evolving Agent Architecture
量子位 · wechat · 2026-08-14
DeepSeek and Peking University published a paper detailing Cordis, the core framework behind DeepSeek Harness, addressing spatiotemporal composability.
Traditional plugin systems often require full process restarts upon unloading, which disrupts AI agents. Cordis introduces "revertible effects" for state rollback (time dimension) and "reactive coeffects" for dynamic dependency resolution (space dimension).
Validated over 4 years in the Koishi chatbot framework with 4,000+ plugins, this architecture provides the foundational infrastructure for self-evolving agents to dynamically manage their own tools.
More from coding & agent
- Amp Code's Orbs Feature Runs Test Suites Remotely to Save Local Compute — bytebot · 2026-08-14
- DeepSeek Harness Source Code Analysis: Prompt as a Runtime — dotey · 2026-08-14
- Developer Builds AI Dungeon Master with DeepSeek: 1000+ Turns for Under $2 — zacurryy · 2026-08-14
- DeepSeek Harness: ~20% of Project Commits Generated by OpenAI Codex — steipete · 2026-08-14
- PathSim: An Open-Source Python Framework for Dynamical System Simulation — tom_doerr · 2026-08-14
- Grok Build Launches Workflows: Orchestrate Hundreds of Agents in Parallel — elonmusk · 2026-08-14