Amazon paper: multi-agent setups burn 1.4-4.7x tokens for small gains; single agent is the default
rohanpaul_ai · x · 2026-11-11
An Amazon paper ran a controlled evaluation of agent architectures—single agent, skill-augmented agent, orchestrator-with-workers, and swarm across 8 variants on Claude Opus 4.8 over 13 enterprise SOPs, holding model, tasks, and tools fixed.
Key findings:
- No universal best architecture: task success ranges from under 10% to near 100% by SOP difficulty; differences between architectures are statistically inseparable, with complex branching SOPs the only plausible case for added structure.
- Added cost must be justified by gains: every multi-agent or skill setup spends 1.4–4.7x the tokens and 2.6–5.8x the latency of the single agent for usually small gains. High-quality skills lift success to 73.1% from 67.4%; degraded skills fall below baseline.
- Implementation quality outweighs architecture choice: a first multi-agent build lost 58.2% to 67.4%, but a targeted redesign hit 75.6%—the only arm to significantly beat baseline, at the highest token cost.
Advice: start with a single agent, measure success per workflow, and fix implementation bugs before paying for complexity.
Related event: Amazon Paper: Multi-Agent Architectures Cost More, Gain Little(2 posts)→
More from coding & agent
- FluentCRM MCP Server lets you manage marketing automation from Cursor or Claude — modelcontextprotocol · 2026-11-11
- Elvis Saravia: agents will subsume traditional software — and can even cut code bloat — omarsar0 · 2026-11-11
- Using a third-party inference provider with your coding harness hands it shell access to your machine — TheZachMueller · 2026-11-11
- Splash One: flight combat sandbox built with Opus 5.5 in PlayCanvas — willeastcott · 2026-11-11
- Next.js 16.5 delivers 3.3x faster builds for a leading AI lab — cramforce · 2026-11-11
- Plurai launches free agent simulation testing inside Claude Code and Cursor — PrajwalTomar_ · 2026-11-11