Block Berd and dsh Released — The Era of Assembling AI Agents

The hottest topic among AI developers these days is "agents assembled to my preference." Lightweight execution environments that allow you to easily swap out AI "brains," tools, and even decision-making structures like components are gaining attention. DeepSeek's preview release of DeepSeek Harness (dsh) and Block's Berd, led by Jack Dorsey, are at the center of this exciting trend.

DeepSeek Harness — An Engine Where Even the Loop Can Be Swapped Like a Plugin

DeepSeek Harness (dsh), released by DeepSeek as a developer preview, is a local agent execution engine built on the unique philosophy that "the model is the soul, and the harness is the body." While previous agent frameworks made it quite cumbersome to change established operational methods, dsh is designed so that every component forming an agent can be easily swapped out.

Powered by the open-source plugin framework Cordis, it allows you to swap not just models or tools, but even the "reasoning loop" responsible for the agent's thoughts and plans as if they were plugins. It essentially keeps the core structure intact while allowing you to reassemble the agent's judgment structure and behavioral rules to suit any situation.

Installation and execution are also very intuitive. Without the need for complex dependency installations, just type npx @deepseek-ai/dsh web into the terminal to launch the local web interface immediately. You can assemble and test your own custom agent quickly and lightly in your browser, without having to manually edit complex backend code.

Block Berd — A Dedicated Agent Workspace Running on Your Computer

If DeepSeek Harness is the powerful, invisible engine driving agents, Block's Berd, unveiled by Jack Dorsey's team, is the smart workspace where those agents actually operate. Developed with Tauri and React, this desktop app provides an intuitive dashboard to view and manage various agents running on your computer at a glance.

Previously, running autonomous agent tools like Goose or Claude Code required opening a terminal and dealing with tedious console commands. With Berd, you can register these tools as easily as virtual characters. You can even keep multiple agents active on the screen and have them process different tasks in parallel.

Through an intuitive interface, users can change an agent's role and personality, and easily swap out the connected backend AI model. It provides a reliable control center to build and command your own AI collaboration team right on your computer, without needing to dig through complex commands or configuration files.

Standards and Frameworks for Easier and Lighter Connections

This shift toward modular, assembly-style agent creation is possible because the invisible standards and tools supporting it are evolving intelligently. Among them, the most notable move is the transformation of the Model Context Protocol (MCP).

Recently, MCP moved away from its old, complex persistent connection method and completely overhauled to a stateless request-response structure. Simply put, instead of holding onto connections, it now handles requests and responses lightly on the fly. This opens the way for horizontal scaling of local equipment, similar to how large websites use load balancers to distribute traffic across servers, pushing beyond the limits of local hardware.

Combined with Mastra, a TypeScript-based framework with excellent type safety, it offers developers very practical options. Mastra solves the hassles of memory management and agent execution flows. With lighter standards and practical tools combined, you can now quickly assemble stable agents by connecting necessary parts without having to build massive systems from scratch.

Beyond Prompt Engineering to Modular Infrastructure

Creating AI agents is evolving beyond just refining prompts into building reliable, modular local infrastructure.

As suggested by DeepSeek Harness and Block's Berd, lightweight development methods where you snap necessary parts together like Lego bricks, without complex integration processes, are becoming the mainstream.

If you want to build your own smart agent workspace, why not start assembling it today with some lightweight open-source tools?