GitHub

DeepSeek Harness: Composable Agent Architecture

Plugins, replay, policy, and surfaces make agent behavior inspectable

8 chapters95 min total
Keywords
DeepSeek HarnessCordisPluginsAgentSessionToolsPolicyReplayProfilesCompaction
Overview

This series explains DeepSeek Harness as a runtime architecture for composing model-driven work from plugins rather than one entangled application. It moves from Cordis services, injected readiness, typed events, scopes, and effect-based teardown to the Agent’s turn and step lifecycle, tool and policy seams, append-only Session replay, persistence, compaction, and Web, headless, SDK, and ACP surfaces. The central value is explicit ownership: providers can vary behind contracts, model-visible state can be reconstructed from durable events, and application boundaries can be chosen deliberately. The series also preserves the limits: scopes are trusted same-process routing, the sandbox primarily governs stated file effects, and the developer-preview project has not undergone a security audit or independent comparative evaluation.

Learning outcomes
  • DeepSeek Harness makes composition a runtime concern: plugins own services, events, registrations, and teardown.
  • The Agent loop owns turns and steps, while the Session log remains the durable source for model-visible history.
  • Capability seams make providers replaceable, but compatibility, cancellation, failure behavior, and cache effects remain real costs.
  • Scopes and file-effect sandboxes provide stated controls, not complete isolation for untrusted workloads.
Keywords
DeepSeek HarnessCordisPluginsAgentSessionToolsPolicyReplayProfilesCompaction

Chapters

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans
Ch.1-3 FreeCh.4-8 Premium

FAQ

What is DeepSeek Harness?

DeepSeek Harness is an open-source agent runtime that composes model access, tools, storage, policy, and application surfaces through plugins.

How does DeepSeek Harness make agent behavior replaceable?

It separates service definitions, providers, consumers, and lifecycle events, allowing implementations to change behind explicit contracts while differences remain visible.

How does DeepSeek Harness preserve agent history?

It derives model history from an append-only Session log; live streams and projections remain separate, while compaction changes the future surface without deleting raw events.

Is DeepSeek Harness secure for untrusted workloads?

DeepSeek Harness does not establish that untrusted workloads are secure by default: its scopes are same-process routing and its sandbox primarily governs file effects, while the project remains unaudited developer-preview software.