Deepseek 刚刚发布了史上最强的 Claude Code 工作流

Description

The DeepSeek super agent fixes the biggest problem with running AI agents across multiple Claude Code sessions. We break down the DeepSeek harness, what DeepSeek tested, and how we turned it into a Claude Code workflow that built our features and fixed a bug on its own. Thank you Mobbin for sponsoring this video! Get 20% off Mobbin Pro's annual plan for your first year at https://mobbin.com/ailabs. Community with All Resources: https://ailabspro.io/?v=3DklR44DHnM The Roundup, our daily newsletter covering the AI stories that matter. Join now: https://www.theroundup.so/ Got a software idea but don't have the team to build it? Autometa turns your ideas into production-ready software. Check out https://autometa.dev/yt to see what we can build for you. This one is AI news you can actually use. DeepSeek just released a new way of working with AI agents, and it starts from a problem anyone running several Claude Code sessions has hit: a person can only keep track of about three to five sessions at once before they lose track of which one is doing what. What the DeepSeek super agent is - One agent you hand all your tasks to, instead of switching between sessions yourself - DeepSeek runs theirs as an agent that's always running in the cloud, which they talk to through team chats and GitHub - Two kinds of agents: a router that talks to you and hands out tasks, and workers that each do one task in their own context window - The router doesn't break a task into steps the way a lead agent does with its subagents. Its only job is to keep track of the workers and hand them tasks What DeepSeek tested - A single session that could start subagents against their router and workers setup, five runs each - The slowest task took 15.7 minutes with the router setup and 29.5 minutes with the single session, almost twice as long - The single session kept doing the work itself, so every new message interrupted it, and even when told to use subagents it often didn't - DeepSeek runs its router on a lighter model, and it answered faster as more messages came in: 6.4 seconds at about one message a minute, 1.6 seconds at about 40 How we set it up in Claude Code - A ticket-writing skill, one of the Claude Code skills we built for this, that turns each feature into a ticket with an rd.md (what the feature needs to do) and an eval.md (how the agent checks it's done right) - A board that shows whether each ticket is a draft, ready or done, whether it runs solo (a subagent) or team (agent teams, an experimental feature in Anthropic's Claude Code), and which ticket it's waiting on - A router agent on Opus with high effort, told never to write anything itself, that checks every worker's results against the ticket's eval.md - A watchdog on Sonnet that checks every 10 minutes for agents that are stuck, going in loops or waiting on each other - Worker agents on Sonnet, so the agents the router starts don't burn through your usage limit - A merger that combines each worker's worktree into the main branch, a checker and a tester that test the app in a real browser with Vercel's agent-browser, and a reproducer that chases down bugs How to use Claude Code this way - Open Claude Code as the router with the agent flag, set the permission mode to auto, and give it a ticket number - Mark the features you want built as ready, and the router hands them out and only comes to you when there's a choice to make - In our run, the checker found a bug that was in the app before any of the tickets, the reproducer fixed it, and the router only told us once it was fixed - Works with any AI coding agent: run the ticket skill in whichever agent you use, and if it doesn't have agent teams, ask it to use subagents for team tickets instead What it cost - The router was open for about an hour and built three tickets: the team ticket ran 26 minutes before we stopped it and merged it, and two solo tickets ran at the same time in 11 and 17 minutes - About 230,000 tokens written and about 32.7 million read, most of it agents reading parts of the conversation again, which costs way less - More tokens than a single session, but a better app and no more juggling sessions The full DeepSeek super agent setup from this Claude Code workflow is in AI Labs Pro. 00:00 Intro 00:39 The Problem 01:34 The Setup 02:42 The Test 03:32 Tickets 06:49 Sponsor 07:44 The Agents 10:07 Building 12:42 Results Hashtags: #ClaudeCode #ClaudeAI #aiAgents #Anthropic #Claude2 #howToUseClaudeCode #ClaudeCodeSkills #ClaudeCodeWorkflow