Introduction

The core philosophy of DeepSeek Harness (DSH) is a plugin-based architecture. When building complex agentic systems, a common challenge for developers is how to validate different Agent preset combinations in a controlled environment. dsh-agent-evolution is a public DeepSeek Harness bundle designed specifically for this kind of experiment.

It provides a set of runtime tools for assembling new Agents from explicitly configured DSH presets and executing tasks to collect evidence for later evaluation. The bundle provides experiment primitives, not an automated evolution loop.

Plugin Overview

This plugin is maintained by FriendsHL and licensed under MIT. It is a DSH bundle whose core includes the runtime plugin agent-experiment-runner.

DSH is currently in developer preview, so a specific commit should be locked when using this plugin to ensure stability.

Core Functionality

The plugin provides the following three core tools for controlling the experiment workflow:

  • agent_experiment_list_presets: Lists the presets currently authorized in the deployment.
  • agent_experiment_run: Runs a single task using a single preset.
  • agent_experiment_compare: Runs a baseline preset and a candidate preset sequentially and returns two records without automatically selecting a winner.

The experiment runner tool supports the following limits:
* maxDepth: Limits the hierarchical depth of sub-agents; default is 3.
* maxTokens: Optional output token limit; by default, it inherits from the parent.

In addition, the plugin supports an allowedPresets whitelist mechanism. If it is not configured, only presets from roots configured with trust: system are allowed by default. Once a whitelist is configured, the plugin resolves all specified preset IDs at load time and resolves them again before execution to ensure validity.

Installation and Configuration

Installing this plugin requires the --profile web argument. Because it is in developer preview, it is recommended to install from a specified GitHub commit.

Install from GitHub:

dsh plugin --profile web add github:FriendsHL/dsh-agent-evolution#<commit>

Install for local development:

dsh plugin --profile web add link:/absolute/path/to/dsh-agent-evolution

After installing or modifying configuration, restart the running DSH process so the Loader can reassemble the profile.

Configuration Items:

Config Default Purpose
maxDepth 3 Absolute delegation-depth limit for experiment sub-agents.
maxTokens inherited Optional positive output token limit when not overridden by the tool call.
allowedPresets system-trust roster Non-empty whitelist of preset IDs. If not configured, only roots with trust: system are allowed.
listToolName agent_experiment_list_presets Model-visible discovery tool name.
runToolName agent_experiment_run Model-visible single-run tool name.
compareToolName agent_experiment_compare Model-visible comparison tool name.

Behavior and Considerations

The plugin has the following characteristics in its experiment behavior and system interactions:

  1. No automatic scoring or selection: The plugin itself does not score outputs, select a winner, edit presets or plugins, or automatically publish changes. Evaluation, failure attribution, candidate drafting, promotion, or rollback are separate, deferred capabilities.
  2. Log authority: DSH Session logs are the authoritative record. This bundle does not create a second transcript or evaluation store. Callers should inspect the returned persistent session to obtain the full prompt, tool calls, results, and outputs.
  3. SubAgent visibility: SubAgents produced by experiments are ordinary lineage-tracked Agents. Their session headers record parentSession, delegationDepth, and agentPreset, but they do not appear in the SubAgent catalog, continuation, or control APIs.
  4. Dependency requirements: The plugin depends on multiple DeepSeek Harness packages as peer dependencies; installation should ensure those dependencies satisfy the version requirements.

Use Cases

This plugin is suitable for developers and operators who need to manually validate Agent composition strategies. When you need to compare different prompt combinations, model configurations, or tool-calling chains, you can use the agent_experiment_run or agent_experiment_compare tools for controlled testing.

Conclusion

dsh-agent-evolution provides a standardized experimentation framework for the DeepSeek Harness ecosystem. Through strict configuration and explicit runtime tools, it enables developers to safely test Agent combinations.

Plugin home: https://www.skillhub.cn/plugins/FriendsHL/dsh-agent-evolution
Code repository: https://github.com/FriendsHL/dsh-agent-evolution