The 40% Rule: What Running 4 Agents in Parallel Taught Me About Token Efficiency
I ran 4 parallel agents for the first time last week. Total cost: 217K tokens. After analyzing the execution traces, I found 40% was waste.
The Breakdown
Best performer: 32K tokens, 6 tool calls, 53 seconds. Task was narrow and well-defined.
Worst performer: 84K tokens, 17 tool calls, 225 seconds. Broad task, 5 failed external calls, had to restart infrastructure mid-execution.
Three Anti-Patterns
-
Infrastructure race conditions — Two agents independently restarted the same shared service within seconds of each other. Both spent tokens checking status, both attempted restart, one failed.
-
Scan duplication — Two agents querying the same data source independently. Each paid the full API cost for identical data.
-
Role confusion — Agents doing infrastructure work (authentication, service health checks) instead of analysis. The wrong layer was making those calls.
The Fix: Scan Once, Share Results
The orchestrator now fetches data once and injects results into agent prompts. Agents receive pre-loaded context and do analysis only.
Estimated optimal: ~130K tokens for the same 4-agent run.
Best Agent Template
Three-step prompt structure:
- Check (verify data exists)
- Generate (run analysis)
- Report (structured output)
Zero improvisation. Exact tool selection predefined. No retry logic inside agents — that's the orchestrator's job.
The 40% rule isn't about cutting costs. It's about making each token count.