AT-301a · Module 1

Token Economics

4 min read

Every agent in a team maintains its own context window — that is a minimum of 10,000-17,000 tokens per agent just for system overhead. A 5-agent team starts at 50,000-85,000 tokens before any work begins. Add inter-agent communication, file reads, and debate rounds, and a single orchestration can easily consume 500,000+ tokens.

One user ran an OpenClaw security audit with agent teams — 10 scanner agents, 4 documentation agents, 2 debate agents — and spent approximately $80 for 15 minutes of work. Powerful, but expensive. Understanding the cost model prevents sticker shock and helps you design efficient teams.

Do This

  • Use Sonnet for specialist agents, reserve Opus for the lead — 3-5x cost savings
  • Set explicit budgets per session: "This orchestration should cost under $10"
  • Monitor spend with /cost during sessions — intervene if accelerating
  • Keep teams to 3-5 agents — the sweet spot for most tasks

Avoid This

  • Run all agents on Opus when Sonnet handles specialist tasks just fine
  • Leave background agents running after they deliver results — idle agents burn tokens
  • Scale to 10+ agents without computing the probability and cost implications first
  • Forget that a sub-agent doing the same work costs 5-10x less than a team agent

The critical distinction: sub-agents consume 50,000-100,000 tokens internally but return only a 2,000-token summary. That is a 50x compression ratio. Agent teams maintain full context for every member. If your tasks are independent and do not need inter-agent communication, sub-agents are dramatically cheaper. Reserve agent teams for tasks that genuinely require debate, shared iteration, or consensus building.