Research-focused agent item on why token usage in coding agents varies dramatically and does not reliably map to accuracy.
This is one of the most concrete sources because it moves the discussion from vibes to measured agent behavior on coding work.
The paper supports a core site thesis: agent token burn can vary sharply, and more tokens do not automatically mean better results.
A primary source for agent-burn pages, especially when explaining why budgets, traces, and evals matter for long-running coding agents.
langchain-ai/langgraph
A framework for building resilient stateful agents with explicit graphs, persistence, human-in-the-loop flows, and controllable execution.
promptfoo/promptfoo
A CLI and CI workflow for testing prompts, agents, and RAG systems across models, with evals and red-team style checks.
stanfordnlp/dspy
A framework for programming and optimizing language-model pipelines rather than hand-tuning one prompt at a time.
Mahnoor Faisal opened a new Claude Code session, ran /context, and found 51,400 tokens already loaded. Disabling four test plugins and auto-memory got the starting context down to roughly 41,400 before any real prompt.
XDA's Mahnoor Faisal ran five coding jobs on Sonnet 5 twice from an identical starting codebase, changing only the effort level. High spent about 26,000 output tokens; Medium finished the same work on roughly 14,300.
Climate scientist Zeke Hausfather metered his own Claude Code habit: 1,138 typed prompts fanned out to more than 14,000 model calls and 3.2 billion tokens in eight weeks, drawing roughly 170 kWh of data-center electricity.