Why DSH, Deepagents and Pi Are Converging on the Same Harness Architecture

The August 25, 2026 paper “The Empire, Long Divided, Must Unite” conducts a source-level study of three open coding-agent harnesses: LangChain deepagents, Earendil Works' pi and DeepSeek's dsh.

They begin from different philosophies—batteries included, radical minimalism and “everything is a plugin”—yet the paper finds that they converge toward a similar architectural middle form.

Five points of convergence

1. The agent loop becomes infrastructure

Model calls, tool execution, result insertion and continuation form the loop. As projects mature, the loop becomes more standardized and differentiation moves to context, policy and extension layers around it.

2. Append-only, replayable session records

Long tasks require more than the final message. All three systems move toward event histories, state recovery and replay. DSH's Session Log and Trajectory express this direction.

3. Model quirks become data or adapters

Models differ in tool schemas, reasoning modes, context limits and message formats. Mature harnesses isolate these differences in providers, adapters or configuration instead of scattering them through the core loop.

4. Progressive context disclosure

Putting every file, tool description and historical turn into one prompt is expensive and unstable. The harnesses increasingly load, compress, summarize and recover context so the model sees what the current step needs.

5. Explicit extension seams

Deepagents uses middleware and storage interfaces, pi keeps a small core with visible extensions, and DSH composes Cordis plugins through configuration. The mechanisms differ, but customization is separated from core execution.

Important differences remain

Convergence does not make the products interchangeable. Deepagents emphasizes integrated application capabilities, pi emphasizes minimalism and transparency, and DSH emphasizes plugin composition, trajectories and runtime recomposition.

The missing dimension: external verifiability

The paper argues that none fully solves how an outside party can verify an execution record without trusting the runtime that produced it. Append-only logs help debugging, but are not automatically tamper-evident receipts.

Finance, compliance, reproducible research and critical infrastructure may need signed events, content hashes, provenance and independent verification.

What this means for DSH

The paper places DSH in a cross-harness source study rather than treating it only as “DeepSeek's coding agent.” Its durable advantages will depend on stable extension seams, long-running session reliability, provider adaptation, context quality and eventually verifiable execution—not plugin count alone.

Practical guidance

• Disclose the model, harness, plugin composition and exact versions in comparisons.
• Regression-test session recovery and context compaction.
• Keep provider-specific behavior out of an agent's business logic.
• Give plugins explicit permissions, lifecycles and observability.
• Distinguish an available log from independently verifiable evidence.

Related reading

• JIT-Agent and just-in-time harness generation
• DeepSeek Harness security and prompt-injection research
• DSH agent presets versus plugins

FAQ

Is this an official DeepSeek paper?

No. It is an independent study of pinned open-source revisions, one of which is DeepSeek dsh.

Does convergence make the three harnesses interchangeable?

No. They share architectural directions but retain different APIs, ecosystems, defaults, interfaces and operating assumptions.