Stop Babysitting Claude Code: Copy DeepSeek's Super-Agent

Summary

The video argues that the main limit on running many coding agents is often human attention, not model speed. A single chat can be interrupted by incoming messages, while separate chats require someone to track them and copy information between them.

DeepSeek’s lead-and-worker design

DeepSeek’s approach separates coordination from implementation:

• A lead routes incoming tasks but is not equipped to do the coding itself.
• Each task goes to its own worker, which handles the work and keeps its logs and context.
• The lead can return to the user or dispatch another task without waiting for the worker.

This is stricter than merely instructing a model to use subagents “when appropriate.” In DeepSeek’s scheduling test, that general instruction produced no delegation in five runs. A more forceful instruction produced delegation in four of five runs; a dedicated lead-worker arrangement delegated in all five.

DeepSeek’s timing results suggested modest gains on typical tasks and larger gains on slower ones: about 20 minutes for the general instruction, 14 minutes for the stronger instruction, and just over 12 minutes for the lead-worker setup. The slowest tasks fell from roughly 30 and 29 minutes to about 16. These were scheduling tests using a dummy tool—not coding-quality benchmarks—and DeepSeek described the results as indicative.

Routing, memory, and reporting

The design also uses routing rules so that:

• Unapproved or irrelevant bot messages can be discarded before reaching a model.
• New issues go to the lead for assignment.
• Follow-up comments on a task can go directly to its worker rather than passing through the lead.

Direct routing can reduce repeated context and token use, though the lead may need to retrieve the relevant history later. DeepSeek’s “Chain of Recall” handles that by linking saved items with readable IDs, letting the lead follow references from a message to a thread, worker, and report.

Workers are meant to send concise conclusions and blockers—roughly 100 words in the example—with detailed logs and evidence kept on the pull request. Because the lead may not inspect files or logs directly, the report needs to include enough evidence to evaluate the result.

The video’s Claude Code experiment

The creator tested three approaches on five real SymPy bugs, alongside 15 ordinary messages over 20 minutes. Each setup was run once, using Opus 5.5 at high effort and Claude Code 2.1.314:

• Stock setup: delegated 0 of 5 bugs.
• Stronger delegation instruction: delegated 4 of 5.
• Lead without coding tools: delegated 5 of 5.

The test is suggestive, not a ranking: there was only one run per setup. The stock lead reported all five bugs fixed, but none of the five test groups fully passed. The no-hands setup passed one; some tests may have checked more than the requested changes.

Removing the lead’s ability to code did not, by itself, make it responsive. At the ten-minute mark, it took about six minutes to answer which requests were still open, versus about five and a half for stock, because it was waiting on a worker. The creator then added an instruction to dispatch work in the background. On a tiny test repository, the lead answered in about a second and a half while its worker continued running.

The lead’s token use was reported as about 8 million for stock, 2.5 million with the stronger instruction, and 400,000 with the no-hands setup. The creator cautions that nobody cleaned up the code and that a lead saying “done” is not proof of success.

Permissions, safety, and limitations

In Claude Code, background-worker approval prompts can still appear in the main session, requiring the user to approve actions. The lead cannot approve on the user’s behalf. Choosing “yes, and don’t ask again” reduces interruptions but changes the permission tradeoff.

The creator warns against installing DeepSeek’s public prototype on a normal account because it runs with full access as the OS user. The prototype also differs from the article’s ideal: it tells the lead not to do work but still gives it shell tools. DeepSeek proposed a separate watchdog to check for genuine progress, but the video says that feature was not in the prototype. Bug reproduction and drafting while people discuss the issue were also described as off by default.

The approach does not solve code review or conflicting edits. The creator recommends judging test results, artifacts, and what the pull request produces—not trusting a worker’s claim or trying to review every line. Changes still need human evaluation and approval.

Products and practical guide

The video notes that Cursor Projects and Claude Code Projects also use coordinator-style sessions. Both were described as beta and cloud-based by default, with local execution available for some tasks.

For people using Claude Code tabs on their own machine, or without Projects, the suggested setup is:

1. Create two agent definitions: a lead with dispatch and messaging tools, and a worker with coding tools.
2. Tell the lead to send work to workers in the background and not wait for them; configure the worker for background execution.
3. Add a quiet trigger—for example, report failed CI runs with evidence and otherwise stay silent. The creator says this trigger was not tested and should be tried manually before scheduling.
4. Ask workers for short conclusions and blockers, leaving detailed output on the pull request.

The setup copies the division of responsibility, but not DeepSeek’s full routing across channels. The central takeaway is to reduce the user’s role as scheduler while still reviewing evidence and deciding what gets merged.

Main speakers and sources

• The Hype Kevin: Narrator and author of the Claude Code/SymPy test and setup guidance.
• DeepSeek and Yujia Qiao: Sources of the lead-worker design, scheduling experiments, prototype, and “Chain of Recall” concept.
• SymPy maintainers: Provided the bug fixes and tests used in the creator’s experiment.
• Other sources cited: Claude Code documentation; Cursor’s changelog and blog; public comments or posts by Boris Cherny, Theo, and other users; and comments on review practices attributed to Simon Willison and Factory.