Track cost per finished task, not token price: the agent metric that matters
arpit_bhayani · x · 2026-11-16
Arpit Bhayani argues builders obsess over token pricing while ignoring how fast the agent loop converges. Every tool call resends conversation history; prompt caching lowers but doesn't eliminate cost. Failed attempts add turns with longer histories, so trimming context or downgrading models can backfire — you save cents per call but pay for extra correction turns. The metric to track is cost per finished task, including retries and wasted turns; reduce loop iterations via clearer instructions and more reliable tools.
More from coding & agent
- Swarms' AgentRearrange orchestrates multi-agent pipelines via a single flow string — KyeGomezB · 2026-11-16
- Mapping 11 weeks of coding agent sessions (8,000+ messages) as a river — dbreunig · 2026-11-16
- Microsoft open-sources TokenEconomizer, cutting coding agent costs 18–31% — aigclink · 2026-11-16
- Stanford CS329Z Lecture 7: How AI Agent Memory Works — MemGPT & Memory Types — zainhas · 2026-11-16
- Qwen4 Max lands in OpenCode Go, Alibaba Qwen team confirms — Alibaba_Qwen · 2026-11-16
- Agentic harnesses hit a plateau: "the ceiling on intelligence is the model" — alexisgallagher · 2026-11-16