DeepSeek Agent Harness In-Depth Review: The "Everything Is a Plugin" Paradigm for AI-Powered Development

An AI Agent Toolkit Where "Everything Is a Plugin"

Among the growing landscape of AI coding agent tools, DeepSeek's Agent Harness stands out in a compelling way. Ewan, a senior software engineer with 20 years of engineering leadership experience, calls it "one of the most interesting AI tools I've seen in a long time."

Its core philosophy can be summed up in one line: everything is a plugin. From low-level tools and agent loops to the entire user interface (UI), nearly every component can be extended and customized. This stands in sharp contrast to Claude Code and OpenAI Codex — both are powerful, but deeply customizing their behavior or interface typically requires building a separate application on top of them. Native-level customization is largely off the table.

For developers juggling five or six projects and fifty to sixty tasks simultaneously, this level of customizability is genuinely transformative.

A Web-Based Interface: Solving the Multi-Tasking Problem in the Terminal

Installing Agent Harness is straightforward — just copy a single command into your terminal to get it running. The service defaults to a local port, and you open it in the browser from there.

Interestingly, it runs in a web browser, not as a standalone app or inside a terminal. If you've used Codex, the interface will feel familiar. Ewan admits that after burning through his Codex subscription quota, he was forced back to using Claude Code in the terminal — and found the experience quite painful.

His take is worth noting: despite years of development work and daily use of the terminal and tmux multiplexers, the terminal just isn't the ideal environment for AI-assisted coding. Agent Harness's web interface supports running multiple projects and sessions simultaneously with smooth navigation between them — effectively solving the "tracking bottleneck" that comes with heavy multitasking.

Flexible Permissions and Run Mode Settings

The interface offers a rich set of configuration options:

• Permission settings: You can restrict the agent to operate only within the current folder, or grant full system access.
• Run modes: Standard mode covers most use cases; Minimal mode provides only Bash and a string-replace editor; PTC (Programmatic Tool Calling) mode lets the agent write code to customize its own workflow; and Creative mode is for interactively building plugins and extensions.

Deep Customizability: From Model Providers to Agent Loops

First-time login requires a DeepSeek API key, but you can freely switch model providers. The toolkit supports GitHub Copilot, DeepSeek itself, OpenCode, OpenRouter, and more — even locally running AI models.

Plugin configuration is the heart of the toolkit. You can adjust virtually every aspect of the agent — including tools, sub-agents, and even the agent loop itself — as well as manage the number of parallel tool calls. These capabilities are typically unavailable in other toolkits. Each plugin can be individually enabled or disabled, and users can create custom toolkits and agent presets with only the tools they actually need.

The most impressive feature is the custom UI capability. Ewan built a custom dashboard in about 30 minutes to track all running workloads across the agent toolkit: how many steps each task has taken, how much of the context window is being used, which sub-agents have been invoked, token consumption, how many agents are running, and how many tasks are queued — all at a glance. This kind of visibility is nearly impossible to achieve natively with Claude Code or Codex.

Orchestrator Mode: One Agent Commanding Many

One of the most practical use cases for Agent Harness is running it as an orchestrator for multiple coding agents.

By enabling a specific mode, users can delegate tasks to locally running Claude Code and Codex instances via tool calls. In other words, you can use DeepSeek to drive the orchestration while routing actual coding work to your Claude or Codex subscription quotas.

Ewan demonstrated this live: he asked the agent to review a codebase and identify bugs. The agent created four bug-fixing tasks and one improvement task, correctly labeled each one, and attached detailed descriptions of each issue — including code locations and suggested fixes. He then used orchestrator mode to run all four bug fixes in parallel across separate sessions, while a fifth agent session monitored progress across all four.

One real-world gotcha worth mentioning: despite the agent having RPC call capabilities to create sessions and send prompts, it initially "didn't know it had those tools." Ewan had to repeatedly prompt it and point out its own capabilities — a reminder that current AI agent systems still have meaningful blind spots around tool awareness.

The toolkit also includes a trace view similar to LangFuse, which lets you inspect every tool call the agent made, search audit history, and diagnose why an agent is running slowly.

Safety and Cost: Two Unavoidable Realities

Security Risks in the Plugin Ecosystem

Ewan places particular emphasis on security. Any code installed from the internet can potentially contain malicious payloads — and this is especially dangerous in AI agent plugins, because plugins have access to the agent, and the agent has access to your credentials and your entire machine. His recommendation: always carefully review the source code of any third-party community plugins before installing them.

Pay-Per-Use API Billing vs. Subscriptions: Running the Numbers

On the cost side, Ewan shared real usage data: over two days of roughly 20% programming time, he spent about $2. At full-time usage, that would translate to around $6–$7 over two days.

His conclusion is pointed: API key-based pay-per-use billing is not particularly developer-friendly. Monthly bills are unpredictable and can easily exceed expectations. By contrast, subscription plans offer both savings and predictable monthly spending. He views OpenAI's subscription quota as the most generous option, which is why he'll continue using Codex on a subscription basis for cost reasons.

As for running models locally, Ewan also did the math: getting local AI running properly requires thousands of dollars in high-performance hardware, and the results often still fall short of cloud-hosted models — making it economically questionable. It's essentially "paying more for a worse model."

Conclusion: An AI Development Hub Built for Extensibility

Taken as a whole, what makes Agent Harness compelling isn't any single feature — it's the depth of its extensibility and customizability.

Ewan mentioned he's building an application called AgentOS on top of this toolkit to assign and track tasks for both himself and his agents, with support for sub-task delegation, cloud/VM runners, and agent inbox communication. He does most of his work with the DeepSeek V4 Flash model, which he describes as "very fast, relatively cheap, and capable" — a practical complement to slower Frontier models like Codex.

For developers constantly context-switching across multiple projects and tasks, Agent Harness offers a new kind of workflow: a programmable, visual, multi-agent-capable hub for managing an increasingly complex AI-assisted development pipeline. That may be its greatest value proposition.