Deepseek Just Released The Greatest Claude Code Workflow Ever

Summary

DeepSeek’s “superagent” workflow is designed to reduce the overhead of managing several coding-agent sessions. Instead of switching between sessions, the user gives tasks to a single router agent, which assigns work to separate worker/executor agents.

How the System Works

• Router: Communicates with the user, tracks work, and assigns tickets. Unlike a traditional lead agent, it is instructed not to break tasks into smaller steps or write code; its focus is coordination.
• Executors: Work on individual tasks in separate context windows, helping keep task details isolated.
• Tickets: Before work begins, requests are clarified and turned into detailed specifications. Each ticket includes requirements (rd.md), evaluation checks (eval.md), and status information on a shared board.
• Scheduling: Tickets can be marked for independent, parallel work or made dependent on other tickets. “Team” tickets allow agents to coordinate; “solo” tickets are handled independently.
• Monitoring and review: A background watchdog or sentinel checks whether agents are stalled or going off track. The router compares completed work with the ticket’s evaluation criteria and can send it back for corrections.
• Integration and testing: A combiner merges work from separate working trees. A verifier checks the result, while browser-based testing and reproduction agents help identify and investigate bugs.

Findings and Claude Code Implementation

DeepSeek compared a single session that could delegate work with a router-and-executor setup, running each approach five times. In their reported results, the slowest task took 15.7 minutes with the router system versus 29.5 minutes with a single session. The video attributes the difference to the single session continuing to do work itself and being interrupted by user messages, while the router can stay focused on coordination.

The video’s presenters adapted the approach for Claude Code. They created tickets from a product requirements document, then ran a router that assigned ready tickets to agents. In their example, a team ticket used three agents, and two later tickets ran in parallel. The router also asked the user to make a decision when needed, then passed the choice to the relevant agent. The presenters say the workflow created features and helped find and fix a bug, while reducing the need to track each agent manually.

The trade-off is higher token usage and cost: each worker has its own context window. The presenters used Opus for the router and Sonnet for executors, reasoning that detailed tickets made the less costly executor model sufficient. Their Claude Code setup ran locally, unlike DeepSeek’s persistent cloud-based system. The video mentions remote sessions or the open-source Paseo app as ways to monitor agents from other devices.

Guide and Product Information

• The setup guide covers ticket writing, agent roles, Claude Code permissions, parallel execution, merging, and verification.
• The video’s sponsor, Mobbin, is presented as a source of real product-screen references for coding agents. The presenters compare a generic AI-generated paywall with a version designed using Mobbin references and say the latter looked more like a real product.
• The presenters say their configuration files are available through the AI Labs Pro community.

Main Speakers and Sources

• AI Labs presenters/narrator: Explain and demonstrate the workflow in Claude Code.
• DeepSeek: Source of the superagent design and the reported timing comparison.
• Mobbin: Sponsor featured in a product-design example.