{"version":"1.0","type":"card","id":"8ad57de9-2876-4eee-aef2-348d7d974270","url":"https://stacklist.com/card/8ad57de9-2876-4eee-aef2-348d7d974270","title":"How to Cut Agent Tokens by 2.7x","source_url":"https://x.com/_avichawla/status/2091804330118861239?s=12","note":"This page discusses a method to significantly reduce agent tokens by 2.7 times using an open harness. It provides insights and practical tips for implementing this technique effectively.","image":{"url":"https://ucarecdn.com/5a2fa589-6b5e-4d7f-a304-62938dcde4ac/","alt":"How to Cut Agent Tokens by 2.7x","width":1200,"height":480},"stack":{"id":"1659549d-373d-4391-ba12-5a14d40c19ed","title":"Harness Management","url":"https://stacklist.com/c/technology/stack/1659549d-373d-4391-ba12-5a14d40c19ed"},"created_at":"2026-08-26T11:46:14.470Z","updated_at":null,"aco":{"summary":"Avi Chawla explains how to reduce agent token consumption by 2.7x through optimized harness architecture, focusing on context management, tool execution, and model call frequency. The analysis demonstrates that agent cost overruns stem from runtime harness decisions rather than model performance, with strategies like prompt caching and context flattening providing significant savings.","tags":["agent-optimization","token-efficiency","llm-cost","harness-architecture","prompt-caching","context-management"],"key_entities":[{"name":"Avi Chawla","type":"person","confidence":0.95},{"name":"TrueFoundry","type":"organization","confidence":0.9},{"name":"LangChain","type":"organization","confidence":0.9},{"name":"Anthropic","type":"organization","confidence":0.9},{"name":"OpenAI","type":"organization","confidence":0.9},{"name":"TrueForge","type":"technology","confidence":0.85},{"name":"Claude Code","type":"technology","confidence":0.85},{"name":"Codex","type":"technology","confidence":0.85},{"name":"prompt-caching","type":"technology","confidence":0.88},{"name":"agent-harness","type":"concept","confidence":0.92},{"name":"context-window","type":"concept","confidence":0.9},{"name":"Terminal Bench 2.0","type":"event","confidence":0.8}],"classification":"analysis","language":"en","confidence":0.85,"provenance":{"model":"claude-haiku-4-5","tool":"@stacklist/be@0.1.0","confidence":0.85,"timestamp":"2026-08-26T11:46:36.743Z"},"token_counts":{"approximate":3571,"cl100k":2947},"content_hash":"sha256:51df52af205761e266574ee79e03933a7c0c1fafd11d0d43e62e0c052db06152","acp_version":"0.2","body_available":true,"body_tokens":3571,"visibility":"public","agent_accessible":true,"status":"final"},"_links":{"self":"/api/public/card/8ad57de9-2876-4eee-aef2-348d7d974270.json","html":"https://stacklist.com/card/8ad57de9-2876-4eee-aef2-348d7d974270","md":"/api/public/card/8ad57de9-2876-4eee-aef2-348d7d974270.md","stack_json":"/api/public/stack/1659549d-373d-4391-ba12-5a14d40c19ed.json"}}