6 Claude agents code 24/7, leaving humans only to approve?
N01ennn used Claude Opus 5.5 and 20 open-source repositories to assemble a 6-agent team. Planning, coding, testing, and review all run in tmux. Meanwhile, Gipp wrote an AI employee job description for Viktor and manages it from the first week with weekly reports and a scorecard. Put the two approaches side by side, and you can see where an AI team is truly difficult.
N01ennn showed off a setup on X: Claude Opus 5.5 plus 20 GitHub repositories, with 6 Claude Code agents running in the same tmux session. The roles are split into lead, planner, two coders, reviewer, and ops. They plan, write code, test, and review, while humans are only responsible for approving PRs.
The video shows N01ennn's workstation, with multiple windows corresponding to lead, planner, coder, reviewer, ops, and monitoring. The tweet lists these repositories in six groups, with total stars marked at 1.2M+. Below, the original links and star counts are retained for verification.
- Core: claude-code 154k★ link
- Core: claude-agent-sdk-python 8.4k★ link
- Core: ruflo 80k★ link
- Core: crewAI 60k★ link
- Core: langgraph 44k★ link
- Employee: wshobson/agents 42k★ link
- Employee: awesome-claude-code-subagents 26k★ link
- Employee: superpowers 307k★ link
- Parallel: claude-squad 8.7k★ link
- Parallel: vibe-kanban 28k★ link
- Memory: mem0 68k★ link
- Memory: graphiti 32k★ link
- Tools: MCP servers 92k★ link
- Tools: github-mcp-server 34k★ link
- Tools: browser-use 119k★ link
- Tools: composio 31k★ link
- Infrastructure: E2B 14.6k★ link
- Infrastructure: temporal 24k★ link
- Infrastructure: claude-code-action 9.7k★ link
- Infrastructure: langfuse 37k★ link
This setup looks like stuffing a software team into a terminal. Claude Code is an agentic coding tool in the terminal that can read codebases, execute routine tasks, and handle git workflows. tmux lets multiple sessions share one screen. MCP connects agents to tools such as files, git, and browsers. Put these together, and you can indeed create a workflow that never stops.
But the questions in the comments are closer to production: someone asked whether the 6 agents share one subscription, and someone reminded that the API bill won't clock out with you. Rohan Bhanot pointed out that when the same model writes code and reviews it, the reviewer tends to let its own mistakes slide. Lakshman Turlapati asked whether the lead would step in after the reviewer rejects the coder twice. These questions are harder to answer than star counts.
Gipp's article offers another approach. He wrote Viktor a job description instead of a single prompt. Viktor lives in Slack or Teams and connects to 9,800+ tools. The public numbers Gipp lists are: 70,000+ teams use it, annualized revenue has reached $50M, about seven months after launch. Viktor's presets run on Claude Opus 5.5, and important actions wait for human confirmation. These numbers come from Viktor's public page; Gipp says he checked five customer pages and setup guides but did not independently verify customer numbers. The article also discloses a collaboration with Viktor. The original article is here.
The key to the job description is that it is checkable: cite a source for every number; external messages are drafts only and are not sent; ask before spending money or changing production; save corrections when corrected; write a self-assessment on Friday, and humans grade it from A to F. Gipp's rule is that a responsibility is kept only if it gets an A or B for two consecutive weeks, and permissions are not relaxed just because of one good Friday. He ran it for a week, gave Viktor a B+, and said it saved about one workday.
The case data is also very specific. Hampton avoided a planned $440K hire and built 12 dashboards and 26 weekly tasks in six weeks; at O8, 19 people on a 20-person team use an AI employee; Element Turf went from zero to 62 workflows in two weeks; AlphaSignal, an 8-person team, built 18 workflows in 67 days without hiring a developer; TWL scheduled 12 workflows in 15 days, handing off about 2 hours of manual work per day. These are Viktor's public cases, and when reading them, remember that they are marketing materials. Gipp also reminds readers that the Hampton, O8, and Element Turf pages describe where things went wrong, so read the failure sections first.
One approach forms a team in the terminal, the other acts as an employee in Slack. The common points are clear: human approval and checkable output. What is scarcest for a multi-agent team is the management interface: who is responsible, how to review, and how mistakes are recorded. Star counts cannot be treated as productivity, and neither can videos. Ordinary users do not need to set up 20 repos first; they can start with one rule: cite a source for every number, and wait for confirmation before external actions. Developers can look at agent team orchestration, but first get the reviewer and the bill under control.
Six agents online 24/7 is only the beginning; the Friday self-assessment and the human scorecard determine whether they can stay.
发布时间: 2026-10-29 07:32