DeepSeek Harness: Everything Is a Plugin
Source: — Turing Post / Attention Span
Related sources:
Summary
The video presents DeepSeek Harness as an open-source agent runtime whose central design claim is that almost every capability above the operating-system level should be a replaceable plugin. This includes the model adapter, tools, skills, sessions, storage, sandbox, agent loop, scheduling and user interface. The model is therefore treated as one service inside a wider, composable system rather than as the whole assistant.
The most distinctive feature discussed is Creator mode. An agent can inspect its running environment, create and test a temporary plugin or agent preset, ask the user to start it, use the new capability, and later remove it. The video stresses that this is not autonomous overnight self-redesign: temporary packages live in process memory, a person still has to start them, and the creator sandbox is not a security boundary.
The underlying Cordis runtime provides plugin mounting and unmounting, service dependencies, event handling and ordered cleanup. The accompanying programming-language-theory paper describes temporal composability (effects should have inverses), spatial composability (dependants shut down before providers), and confluence (after temporary experiments are removed, the settled observable system should correspond to the remaining components rather than to every detour taken). These guarantees depend on correctly implemented inverses, suitable dependency structure and other formal assumptions; irreversible external effects such as a sent email cannot be undone.
The broader argument is that abundant generated code may change the economics and architecture of software. Instead of shipping every possible feature in advance, an agent could generate a task-specific capability when needed, use it and remove it. The video compares this possibility with Smalltalk, Unix’s “everything is a file”, and standardised electrical sockets. It also cautions that DeepSeek Harness is a developer preview and a proposal, not yet a cross-vendor standard.
Alf’s evaluation against Hermes
Overall assessment
Interesting architectural complement, but not an immediate replacement for Hermes. DeepSeek Harness is especially relevant to the question of how agent capabilities should be composed and safely reconfigured at runtime. For the household use case, Hermes is currently the better fit: it already combines durable memory, skills, Obsidian integration, scheduled jobs, messaging channels, profiles and deliberately governed privacy boundaries.
Alf’s assessment is that the strongest near-term value is to monitor Harness as an architectural reference and possibly experiment with it in an isolated sandbox. The household workflow should not be migrated merely because the plugin model is elegant or because the repository is attracting attention.
Where Harness is highly relevant
- Capability composition: Its “everything is a plugin” model closely matches the goal of keeping model choice, tools, skills, storage, workflows and interfaces replaceable.
- Generated task-specific tools: Creator mode points toward an agent being able to build a temporary helper for an unusual research or coding task instead of requiring a permanently installed integration.
- Traceability: The append-only trajectory/session-log concept is valuable for auditable research, debugging, resuming, forking and replaying runs.
- Model and runtime experimentation: Minimal, standard, code and creator modes could support controlled evaluation of different model/tool configurations.
- Open-source leverage: The runtime may be useful for learning from or borrowing ideas without adopting a proprietary coding-agent product.
Decision
Claims and caveats
- The video is commentary, not an independent technical audit; its enthusiasm and GitHub metrics are time-sensitive.
- The official DeepSeek page confirms the developer-preview status, plugin architecture, modes and traceability claims.
- The formal guarantees discussed for Cordis depend on assumptions and correctly implemented plugin cleanup; they do not make arbitrary real-world side effects reversible.
- “End of coding agents” is a headline-level framing, not an established conclusion.
Why it matters:(AI) DeepSeek Harness makes the runtime surrounding an AI model visible as a first-class architectural layer. That is directly relevant to the long-term design of Hermes, specialist agents and the household automation system, but it does not currently outweigh the value of Hermes’s integrations, memory, governance and established vault workflow.
Tags: AI Agents DeepSeek Harness Cordis Hermes SoftwareArchitecture OpenSource