From 100M to 5B tokens a day: multi-agent swarms emerge as a new scaling axis
xeophon · x · 2026-09-27
Florian Brandt details his hands-on experience running multi-agent swarms exclusively on open models:
- His personal token usage grew from under 100M tokens/day six months ago to over 5B/day now, likely exceeding 10B by year-end — a 2-OOM jump
- Drivers: models now run unattended for hours, GPU access fuels experimentation, and multi-agent swarms are in their early innings
- Swarms shine on data work and broad research: one agent finds something and relays findings to the group via agent-to-agent communication; open models handle persistent fleets of (sub-)subagents better than expected
- He argues single-agent systems are hitting wall-clock time limits, making realtime budgets the underexplored scaling axis — and guesses that a certain HF agent's sensitive "64H" IP refers to hours-based budgets, which matter for GPT-6-style swarm behavior since token budgets are ultimately inferior
More from coding & agent
- DHH on AI coding: thinking "architecture" is your fortress? Models are good at that too — tawnniee · 2026-09-27
- Reverse engineering with coding agents: set it up overnight, wake up to source — moyix · 2026-09-27
- ThePrimeagen joins DHH's Omarchy Core to lead agentic QA with Oligarchy harness — mark_k · 2026-09-27
- LLMs don't write better code — they accelerate the glue work of integrating libraries — markjeffrey · 2026-09-27
- Open-source AI agents now delegate tasks to each other: GrokBot + Muse — Saboo_Shubham_ · 2026-09-27
- 'Agentic AI Killed DevOps' Claim Sparks Developer Pushback — emeka_boris · 2026-09-27