Tracing agentic calls: one edit burns 760KB; dropping agent loops cut bill from $400 to under $100
cgouguen · reddit · 2026-09-16
After his API bill topped $400/month, the author traced the actual exchanges between coding harnesses and LLMs — and the agent-vs-single-shot gap was bigger than expected.
Same simple PyQt task (change two files' layout):
- Pi: 3 LLM calls, 760KB JSON — read files, edit, then summarize, three round trips
- Aider: 1 LLM call, 100KB — one pre-assembled prompt with repo map, file contents, and SEARCH/REPLACE format instructions
His take: for most daily tasks you already know which files to edit; you don't need an agent to "discover" them and burn tokens in a loop. Since Aider is unmaintained, he built his own lightweight single-shot harness, Frugaast, cutting his bill from $400+ to under $100/month. He wonders whether Cursor and Copilot hide similar token bloat.
More from coding & agent
- HF Jobs OCR script zoo: 28 models, one command, 7 pages for 7 cents — vanstriendaniel · 2026-09-16
- Creator's 3-step system turns 15 manual podcast steps into 8 AI skills, saving 5 hours a week — petergyang · 2026-09-16
- Friends vibe code a personal chat app with retro Windows-style UI, planning a shareable release — floguo · 2026-09-16
- Cumora: open-source team chat where AI agent teams collaborate with humans — tom_doerr · 2026-09-16
- EDSL to add activation capture, probes and model steering for LLM social simulation research — soumitrashukla9 · 2026-09-16
- NoSpoon agent autonomously generates hundreds of AI microdramas daily — Kyrannio · 2026-09-16