I built a NotebookLM-style app with 7 coding agents in 78 hours. Here's what worked and what broke.
retained text
I recently experimented with running seven coding agents as a team instead of working with a single agent. Six agents handled different parts of a NotebookLM-style application: infrastructure, ingestion, retrieval, frontend, accounts, and shared libraries. The seventh agent was the coordinator. The setup was surprisingly simple: * Linux VM with separate GNU screen sessions over SSH * Each agent had its own Git checkout * Direct agent-to-agent communication via file sockets * Shared decision records stored alongside the code * One head agent coordinating tasks, reviewing PRs, merging, and handling releases * CI checks required before merges, followed by automated deployments and smoke tests I interacted mostly with the head agent through Claude Desktop. It delegated work, coordinated dependencies, and brought questions back to me. **The results after 78 hours of elapsed time:** * 650+ merged PRs * 1,930+ commits * 208 test files * Around 81,000 lines of Python and TypeScript * 300 documented engineering decisions * Roughly 20 to 24 hours of my own involvement But the interesting part wasn't the amount of code. **It was how natural the workflow became.** I could discuss a problem with the coordinator while six other agents continued implementing different components. Of course, things broke. Agents collided on shared files. Frontend changes sometimes arrived before their backend dependencies. Release tooling had bugs. And my manual testing uncovered problems that automated tests had missed. Even the coordinator needed supervision. At one point, another agent correctly refused its request because the required human approval hadn't been granted in that agent's session. My biggest takeaway: multi-agent development works surprisingly well when responsibilities, communication, review gates, and decision history are clearly defined. But I now see a different bottleneck emerging. **Human attention.** I don't want to monitor seven conversations to find three questions that actually need my input. I'd rather have one prioritized inbox for decisions, approvals, and blockers, while the agents handle everything else themselves. I've documented the experiment and published the repository, including the build history and what went wrong. Links in the first comment. Curious how others running multiple coding agents handle coordination and human approvals.
document fetched 2026-10-21 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xas3hu/i_built_a_notebooklmstyle_app_with_7_coding