DeepSeek Harness (DSH) Unveiled: Analyzing DeepSeek V4 Pro’s Model, Protocol, and Runtime for Agents

Overview

On August 13 DeepSeek launched the V4 Pro model and the DSH (DeepSeek Harness) runtime. The update introduces a three‑layer architecture for agents: the model layer (V4 Pro / V4 Flash), the protocol layer (Responses API), and the runtime layer (DSH).

Three‑Layer Map

The model layer handles task understanding, reasoning, and next‑step selection. The Responses API transports multi‑turn inputs, outputs, and tool results between client and model. DSH schedules execution, records session events, and enables replay.

DSH Runtime Details

DSH is open‑source under the MIT license and currently in developer preview (version 0.1.0-rc.6). It can be started locally with npx @deepseek-ai/dsh web, which serves a UI at http://127.0.0.1:3080. The runtime treats every component—model adapters, tool registrations, session logs, and the agent loop—as plugins loaded by the Cordis system, allowing independent replacement and versioning.

Plugin Modes

Standard mode : full tool set for complete agent tasks.

PTC mode : model‑generated code with multi‑tool calls to reduce round‑trips.

Minimal mode : only shell and file‑edit tools for baseline performance.

Creative mode : in‑memory experimentation with Cordis plugins.

Start with Standard mode for full‑task testing; switch to Minimal mode when isolating variables.

Session Logging

DSH records a Session Event Log that appends every tool call, parameters, results, and intermediate states. This log enables replay, branching, audit, and reconstruction of the model’s context for the next turn.

Seams for Replacement

Modules such as the model adapter, tool interface, file system, and subprocess are exposed via replaceable seams. Swapping a seam (e.g., a new model adapter) preserves the agent loop and existing logs.

V4 Pro Model Improvements

Official benchmarks for V4 Pro include:

Terminal Bench 2.1: 87.9 (comprehensive terminal tasks)

DeepSWE: 62.7 (software‑engineering tasks)

Toolathlon‑Verified: 74.1 (tool usage and task completion)

HLE (no‑tool / with‑tool): 42.7 / 60.0 (impact of tool integration)

These scores reflect the specific benchmark setup (DSH minimal mode, particular prompts, and tool definitions) and are not directly comparable to production efficiency.

Responses API Enhancements

The new OpenAI‑compatible Responses API now natively supports Codex. It remains stateless; the client must retain multi‑turn history and handle session persistence, compression, and failure recovery.

Pricing Structure

Pricing (as of August 13) is per 1 M tokens, with separate rates for cache‑hit input, cache‑miss input, and output. Example prices for V4 Pro:

Cache‑hit input: $0.003625

Cache‑miss input: $0.435

Output: $0.87

Peak‑hour rates (09:00‑12:00, 14:00‑18:00 CST) are roughly double non‑peak rates. Non‑peak rates for V4 Pro are $0.022 (cache‑hit), $0.66 (cache‑miss), $1.98 (output).

Cost Considerations

Cache hits depend on request structure; stable system prompts and tool definitions should be placed in the prefix.

Batch jobs can be scheduled in non‑peak windows to reduce cost.

Only looking at input token price is insufficient; output tokens, retries, and reasoning intensity also affect the bill.

Measure your own task set for cache‑hit ratio, tool‑call success, test pass rate, latency, and total cost before production deployment.

Agent Task Flow

A typical agent task proceeds as: Task → V4 Pro → Responses API → DSH Runtime → Tool / File / Test → Verifiable Result . DSH records each round, allowing the system to decide whether to continue, retry, or stop based on verification outcomes.

Practical Recommendations

For initial experiments, directly call deepseek-v4-pro or deepseek-v4-flash. To assess production suitability, pick a task with clear acceptance criteria, run it in Standard mode, then in Minimal mode to isolate variables. Record model version, reasoning intensity, cache‑hit status, tool‑call count, token usage, latency, and test outcome.

Conclusion

DeepSeek’s update raises the model ceiling for tool‑driven tasks, provides a more agent‑friendly protocol, and opens a plugin‑based runtime for deeper inspection and customization. However, real‑world costs depend on token pricing, cache efficiency, scheduling, and the number of agent loops, so thorough measurement is essential before adopting the stack in production.

Prices and model versions may change; always verify the latest information on DeepSeek’s official pages.