DeepSeek Drops Harness: The Modular Coding Agent That Stole 37k Stars
DeepSeek just open-sourced Harness, turning the nightmare of modular AI coding agents into a single one-line command that works across any model provider.
Under an MIT license, the developer preview repository for Harness raked in over 37,500 stars and nearly 3,000 forks on GitHub in its first twenty-four hours.
The system spins up immediately via a single npx command, launching a local Web UI where an agent reads and modifies codebases, executes shell tasks, plans workflows, searches the web, and delegates tasks to subagents without asking for a massive infrastructure setup.
Everything under the hood is powered by the lightweight Cordis microkernel. Instead of locking developers into hardcoded workflows, every core mechanic—from model adapters and sandbox environments to the agentic loop itself—runs as an interchangeable plugin.
The framework comes preconfigured with four operational modes: Standard, Minimal, Create (where the agent inspects memory to generate new modes on the fly), and PTC (Programmatic Tool Calling). PTC enables the model to generate a single TypeScript script that batches multiple tool executions together, drastically reducing the token lag of back-and-forth roundtrips.
Configuration profiles also avoid vendor lock-in completely, supporting plug-and-play integrations with OpenAI, Anthropic, or any custom API-compatible server.
While proprietary giants continue locking agentic workflows behind costly monthly subscriptions, fully open-source modular toolkits running right out of the local terminal demonstrate how quickly centralized developer tools are turning into commoditized building blocks.
Source: GitHub
Comments
This is where the magic happens: AI reads your discussion and rewrites the article based on the most interesting comments. Each strong comment adds points to the meter below. Once the meter is full, the article updates live — no page reload needed.