DeepSeek and Peking University Unveil Spatiotemporal Composability Paper and Cordis Framework for Self-Evolving Agents
DeepSeek and Peking University jointly released the paper "A Programming Paradigm for Spatiotemporal Composability" along with an accompanying test harness, detailing Cordis, the core framework behind DeepSeek Harness. The work tackles the stability of self-evolving agents that continually rewrite their own runtime environment, enabling runtime hot-plugging of components and clean undoing of side effects, and is read as DeepSeek's first substantive attempt at framework-level continual learning.
Confirmed
- Core model: Per @tokenbender, the paper proposes a component-level adaptive composition programming model introducing "adaptive composition/formalization"; components explicitly declare their required inputs and the changes they will produce, and the runtime automatically activates or deactivates dependencies as components join or exit, cleanly undoing side effects.
- Cordis and hot-pluggable architecture: Per @量子位, the paper details the Cordis framework behind DeepSeek Harness, which defines a persistent abstraction layer so tools, memory, skills and other components can be dynamically plugged in and out at runtime; traditional plugin systems such as VSCode usually require restarting the whole host process on uninstall, losing state in self-evolving AI scenarios, and this design addresses dynamic registration and state restoration.
- Companion release: The paper shipped together with a companion test harness, a point relayed by @SimplyAnnisa and others.
Why it matters
- Per the takeaways relayed by @philipvollet, current recursive self-improvement (RSI) discussions focus mostly on model weights or algorithmic optimization, while real agents must keep rewriting their harness — tools, memory, sandboxes; without safeguards, self-modification easily causes irreversible side effects or breaks recovery mechanisms.
- Per the interpretation relayed by @teortaxesTex, the work points to a new tool paradigm for self-evolving agents: unlike minimal setups that only run scripts via bash, evolving agents dynamically create and discard tools, demanding strong dynamic registration, state restoration and side-effect handling; the discussion also touched on TypeScript's position in this new paradigm.
2026-08-13 ~ 2026-08-14 · 6 related posts
- Episode 1: DeepSeek-Harness WebUI Allegedly Leaked Ahead of Launch(2026-08-12, 2 posts)
- Episode 2: DeepSeek Open Sources Plugin-Based Agent Framework Harness v0.1(2026-08-13, 36 posts)
- Episode 3: DeepSeek-Harness Launch Splits Opinion: Architecture Praised, Usability Panned(2026-08-13, 6 posts)
- Episode 4: DeepSeek Harness Hits 100k Stars in 48 Hours(2026-08-13, 6 posts)
- Episode 5: DeepSeek and Peking University Unveil Spatiotemporal Composability Paper and Cordis Framework for Self-Evolving Agents(2026-08-13, 6 posts)
- Episode 6: DeepSeek Harness Defended as Foundational Agent OS Beyond Coding(2026-08-14, 2 posts)
- Episode 7: DeepSeek Harness: ~20% of Commits Were Made by Codex(2026-08-14, 3 posts)
Primary sources
- [source] DeepSeek's New Paper: Introducing an Adaptive Synthesis Programming Model — tokenbender · 2026-08-13
- DeepSeek Tackles Continual Learning with Hot-Swappable Agent Framework — teortaxesTex · 2026-08-14
- Peking Univ & DeepSeek Paper: Dynamic Architecture for Self-Evolving Agents — burny_tech · 2026-08-14
- [source] DeepSeek & Peking University Paper Unveils Self-Evolving Agent Architecture — 量子位 · 2026-08-14
- DeepSeek's Self-Evolving Agent Paper: Dynamic Tool Paradigm and TypeScript's Role — teortaxesTex · 2026-08-14
- Solving Agent Self-Destruction in RSI: A Paradigm for Spatiotemporal Composability — philipvollet · 2026-08-14