Why AI Agents Burn So Many Tokens: Context Loops, Tool Schemas and History
Mks-101 · reddit · 2026-09-17
A developer documents how agent token usage explodes far beyond a simple request: system prompts and agent instructions, tool definitions and schemas, conversation history, tool outputs, retrieved documents, intermediate reasoning loops, and repeatedly re-sent context all stack up. Unlike a plain chatbot, an agent effectively reprocesses surrounding context over and over as it decides what to do next.
The post asks the community:
- What are the biggest sources of token consumption in your agents?
- Which techniques actually help: prompt compression, better tool design, history summarization, limiting agent loops, caching, or using smaller models for certain steps?
More from coding & agent
- ChatGPT Work ran 7 hours, silently failed, and admitted its progress updates overstated completion — Leather-Driver-8158 · 2026-09-17
- mattpocock ships /pr skill: PR bodies optimized for fast human review — mattpocockuk · 2026-09-17
- Dev Runs Fully Autonomous Agent Teams: It Works End-to-End, But Not Reliable At All — MoonL88537 · 2026-09-17
- QOJ publishes list of contest problems where GPT-6 Pro found solutions beating the authors' — teortaxesTex · 2026-09-17
- What Happens After You Tell an AI Agent It's Wrong? Dreamforce Enterprise Lessons — TheTuringPost · 2026-09-17
- GitHub MCP maintainer packs the house at MCPCon with talk "MCP doesn't have a context problem" — marlene_zw · 2026-09-17