agent

Paper: AI agents can spend unpredictably on coding tasks

Research-focused agent item on why token usage in coding agents varies dramatically and does not reliably map to accuracy.

Published 2026-04-28Source: arXiv
arXiv source artwork

Why it matters

This is one of the most concrete sources because it moves the discussion from vibes to measured agent behavior on coding work.

Tokenmaxxing read

The paper supports a core site thesis: agent token burn can vary sharply, and more tokens do not automatically mean better results.

Source takeaway

A primary source for agent-burn pages, especially when explaining why budgets, traces, and evals matter for long-running coding agents.

Topic links

Related projects

Tools that match this angle

#4In spirit
Agents

LangGraph

langchain-ai/langgraph

A framework for building resilient stateful agents with explicit graphs, persistence, human-in-the-loop flows, and controllable execution.

45.2K7.8KMIT
agentsstateworkflows
#5Direct
Evaluation

promptfoo

promptfoo/promptfoo

A CLI and CI workflow for testing prompts, agents, and RAG systems across models, with evals and red-team style checks.

27.5K2.8KMIT
prompt-evalscirag
#6In spirit
Evaluation

DSPy

stanfordnlp/dspy

A framework for programming and optimizing language-model pipelines rather than hand-tuning one prompt at a time.

39.2K3.5KMIT
optimizationprogrammingevals
Related feed

More source-linked context

bestagent.dev source artwork
newsB
news

Claude Code's Code Mode, tested: about 22% cheaper on two of three tasks

bestagent.dev ran three tasks three times each on Claude Code 2.1.325 with Opus 5.5, Code Mode on and off. A feature-with-tests job fell from $0.45 to $0.35 a run; a broken-link check from $0.20 to $0.16. A 20-file edit barely moved.

tokenmaxxingcoding-agentsagents
Read note
XDA source artwork
long-formX
long-form

Claude Code was using 51,000 tokens before I even typed a prompt — I fixed it

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.

coding-agentstoken-consumptiontoken-waste
Read note
XDA source artwork
long-formX
long-form

Dropping Claude Code from High to Medium effort cut output tokens 45%

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.

coding-agentstoken-consumptionai-spend
Read note