DeepSeek Harness Settings & Agent Presets: Which Mode Fits Your Task?

General Settings

DeepSeek Harness's general configuration panel covers five core dimensions: Agent Preset, Permission Policy, Interface Language, Appearance Theme, and Enter Key Behavior When Busy.

Permission Policy

Three incremental levels restrict the agent's file system and terminal command capabilities:

Read Only – Only reads workspace files; cannot modify files or execute terminal commands. Highest security.

Workspace Write – Allows read/write within the current working directory and command execution in that directory (requires user approval). Cannot access files outside the workspace. Recommended for daily development.

Full access – Full file system access; can read/write arbitrary paths and execute terminal commands without approval. Higher security risk.

Enter Key Behavior When Busy

Queue Send (default) – Does not interrupt the current turn; messages enter a FIFO queue and are processed after the current task finishes.

Interrupt Send – Immediately interrupts the ongoing turn for priority handling; used for urgent corrections or direction changes. If the interrupt window has passed, it automatically falls back to queue mode without losing messages. Note: this is a "cooperative interrupt" decided by the host AgentLoop, not a preemptive kill.

Model Configuration

Harness uses a pluggable adapter architecture, natively supporting DeepSeek's official API endpoints while remaining compatible with third‑party model services. Users can add providers (e.g., Zhipu's coding‑plan suite with an API key) and switch models via the conversation header selector. For enterprise private networks or locally deployed models (vLLM, LocalAI, TGI), a Custom Provider option lets you specify OpenAI‑compatible endpoint metadata.

Plugin System

Harness follows an "Everything is a Plugin" philosophy. The current version ships with 165 built‑in plugins that collectively provide the web app's entire functionality—UI rendering, message flow, model invocation, tool execution, file operations, etc. The plugin list shows name, version, status, and description for quick auditing.

Plugin Parameter Configuration

Selected plugins expose runtime parameters:

Terminal Plugin – Command Timeout (seconds) limits single shell command runtime; Harness auto‑terminates on timeout.

Agent Loop Plugin – Parallel Tool Calls controls the maximum number of tools the agent may invoke concurrently in one reasoning cycle. Raising it speeds multi‑tool tasks but must respect downstream API rate limits.

Web Search Plugin – Configures third‑party search API endpoints and credentials (e.g., Bing Search API, Google Custom Search JSON API) to give the agent real‑time retrieval capability.

Agent Presets

An Agent Preset is a complete environment parameter set covering tool set, system prompt, context window strategy, and task orchestration logic. Switching presets changes the agent's working style and capability boundaries within the same framework. Four built‑in modes exist:

4.1 Standard Mode (Default)

Loads the full general‑purpose agent toolkit: file read/write, terminal commands, web search, task planning, Skills, sub‑agents, todo management—about 20+ tools . Comparable to tools like Codex, it handles end‑to‑end development from requirement understanding to code delivery. The 3D racing game demo in the previous article ran in this mode.

Trade‑off : Every inference request encodes all 20+ tool definitions into the system prompt, consuming nearly 10k tokens even for a simple greeting. Users can inspect the full system prompt and context snapshot via the Trace button.

4.2 Minimal Mode

Keeps only two tools: a persistent terminal and a file editor. System prompt reduces to "You are a helpful software engineer assistant." A greeting now uses only ~ 1.2k tokens . The agent behaves as a focused execution engine—no proactive task decomposition, progress reporting, or multi‑path exploration. DeepSeek positions this as the standard environment for model benchmarking ; DeepSeek V4 Pro's public scores were obtained in this mode.

4.3 PTC Mode (Programmatic Tool Calling)

Instead of sequential tool calls, PTC bundles repetitive, parallelizable operations into a single run_code invocation that executes batch logic via internal APIs. Only run_code is exposed, yet it indirectly provides the full standard toolset (file ops, network requests, system commands). Ideal for batch file renaming, dataset preprocessing, automated test suite runs—tasks with clear, linear steps.

4.4 Comparative Benchmark: 3D Design Game Task

The author replicates a 3D game creation task (shared by community expert "Yupi") across three modes:

Standard – 33 minutes, ~5.2M input tokens, high visual fidelity.

Minimal – 23 minutes, significantly lower token usage, but simplified models/textures and lower visual quality.

PTC – 27 minutes, ~1.5M input tokens (under one‑third of Standard), final result similar to Standard.

Guideline : Use Minimal for speed‑sensitive, simple logic; Standard for stability and generality; PTC for multi‑step, clearly defined batch/automation tasks.

4.5 Creative Mode

Builds on Standard's full capabilities and injects meta‑tools that operate on Harness's underlying Cordis plugin system. These allow the agent to query loaded plugins, experiment with new plugin combinations, and generate new preset configuration files at runtime. Users can create custom agents via natural language without writing YAML/JSON.

Example: Replicating Cursor's Debug Mode

Prompt:

基于标准模式创建一个自定义 Agent 预设，名称为「Debug 模式」。要求复刻 Cursor IDE 的 Debug 模式工作方式。必须参考 Cursor 官方文档来设计：https://cursor.com/docs/agent/debug-mode

Set model to DeepSeek‑V4‑Pro with MAX reasoning ; permission to Full access for writing global config.

Harness loads the "Edit Preset" skill, fetches Cursor's docs via terminal, extracts the debug workflow (run tests → analyze logs → insert diagnostic logs → reproduce → root‑cause analysis), and generates two files: agent.cordis.yml – defines toolset, system persona, and debug workflow rules. preset.yml – controls UI display name and description.

The new "Debug Mode" appears in the mode selector for future conversations.

This mirrors AGENTS.md or CLAUDE.md style constraints but with finer granularity, higher automation, and no manual structured‑doc maintenance.

Summary

The article comprehensively analyzes DeepSeek Harness's general configuration and agent preset mechanism. General settings cover permission policies, model integration, and the plugin system. Four built‑in presets are contrasted: Standard (feature‑complete but high context overhead), Minimal (lightweight, benchmark‑oriented), PTC (batch‑oriented efficiency via programmatic tool calling), and Creative (self‑modifying capability enabling natural‑language creation of custom presets such as a Cursor‑style Debug mode). Mastering these preset differences is key to harnessing Harness effectively.