I ran 7 coding agents for 78 hours: 650+ PRs, 81k lines, and what broke

oliver-zehentleitner · reddit · 2026-10-20

A developer shared a full retrospective on building a NotebookLM-style app with 7 coding agents working as a team.

Setup

Linux VM with separate GNU screen sessions; each agent had its own Git checkout
Six agents owned infra, ingestion, retrieval, frontend, accounts, and shared libraries; a seventh acted as coordinator
Direct agent-to-agent communication via file sockets; decision records stored alongside code
CI checks required before merges, then automated deployments and smoke tests
Human interacted mainly with the head agent via Claude Desktop

Results in 78 hours: 650+ merged PRs, 1,930+ commits, 208 test files, 81k lines of Python/TypeScript, 300 documented engineering decisions—with only 20-24 hours of human involvement.

What broke: agents collided on shared files, frontend changes arrived before backend dependencies, release tooling had bugs, and manual testing caught issues automated tests missed. The coordinator needed supervision—at one point an agent correctly refused its request because human approval hadn't been granted in that session.

Biggest takeaway: multi-agent development works surprisingly well when responsibilities, communication, review gates, and decision history are clearly defined. The emerging bottleneck is human attention—he wants one prioritized inbox for decisions and approvals instead of monitoring 7 conversations. Repo and build history are public.

Original post →

I ran 7 coding agents for 78 hours: 650+ PRs, 81k lines, and what broke

More from coding & agent