10,000 AI agents burned his API credits; 500 agents in 50 waves worked

Beneficial-Cow-7408 · reddit · 2026-11-03

A developer built Flow with Claude Code: a live canvas where hundreds of AI agents collaborate on one evolving project.

What went wrong

Running 10,000 agents silently stalled after 4 hours — no code errors, just 10 notifications each from OpenAI and Anthropic: API credits exhausted.
He rebuilt with 500 agents across 50 waves, up to 10 concurrent.

How the system works

Each agent has its own role, instructions, model, reasoning effort, and assigned assets, all editing the same evolving project file with patches landing live on the canvas; each wave builds on the previous compiled result.
Claude, OpenAI, Gemini, and Grok can share one workflow, each assigned tasks by strength.
Also demoed: multiple versions from one starting state, 4,000+ accessible project states with downloadable files, cloud execution after closing the browser, mobile/desktop continuity, Split Screen for two projects, Split File for up to 4 versions of one file.

Lessons

Using Opus-class models for 200+ agents was wasteful. The right split: a frontier model for architecture, integration, and final compilation; lighter models for small additions. He found 100 lighter agents handle focused changes fine, but once a single file hits 25,000 lines a stronger model is still needed to compile and integrate cleanly.

More from coding & agent