World PulseNow
Newsletter

GitHub Slashes Agent Workflow Token Spend up to 62% with Daily Audits and MCP Pruning

InfoQ — AI, ML & Data Engineering•Friday, May 29, 2026 at 8:30:00 AM
PositiveArtificial Intelligence
GitHub Slashes Agent Workflow Token Spend up to 62% with Daily Audits and MCP Pruning
  • What Happened

    GitHub has successfully reduced token costs in its agentic CI workflows by up to 62% through the pruning of unused MCP tools, replacing some MCP calls with gh CLI, and implementing daily audits via “auditor” and “optimizer” agents. This initiative also includes the introduction of a token-usage.jsonl artifact and an Effective Tokens metric to monitor spending and identify regressions.

  • Why It Matters

    This cost-saving measure is significant for GitHub as it seeks to optimize its operational efficiency amid rising expenses associated with AI tools and services, particularly as it prepares for a transition to a usage-based billing model for its Copilot service.

  • The Bigger Picture

    The shift to a metered billing approach for GitHub Copilot, set to take effect in June 2026, reflects broader industry trends towards usage-based pricing models in response to increasing operational costs. This change has sparked discussions about the sustainability of flat-rate subscriptions in the face of escalating AI usage, highlighting the challenges companies face in balancing user demand with economic viability.

— via World Pulse Now AI Editorial System
Read Original

Was this article worth reading? Share it

Continue Readings
Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks
arXiv — cs.CL19 hours ago
Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks
NeutralArtificial Intelligence
The introduction of Claw-SWE-Bench presents a new benchmark for evaluating OpenClaw-style agent harnesses on coding tasks, addressing the challenges of measuring coding abilities in general-purpose agents. This benchmark includes 350 GitHub issue-resolution instances across eight languages, providing a structured environment for assessment.
Read full article
via arXiv — cs.CL
Efficient AI-Inspired Reduction of Feynman Integrals via Tube Seeding
arXiv — cs.LG2 days ago
Efficient AI-Inspired Reduction of Feynman Integrals via Tube Seeding
PositiveArtificial Intelligence
A new machine learning-based seeding strategy for the integration-by-parts reduction of Feynman integrals has been developed, addressing a significant bottleneck in theoretical particle and gravitational-wave physics calculations. This method utilizes a sparse selection of seed integrals, allowing for the reduction of complex multi-loop integrals more efficiently than traditional approaches, which often lead to polynomial growth in complexity.
Read full article
via arXiv — cs.LG
Efficient Imputation for Patch-based Missing Single-cell Data via Cluster-regularized Optimal Transport
arXiv — cs.LG2 days ago
Efficient Imputation for Patch-based Missing Single-cell Data via Cluster-regularized Optimal Transport
PositiveArtificial Intelligence
A new imputation algorithm named CROT (Cluster-Regularized Optimal Transport) has been introduced to address the challenges posed by missing data in single-cell sequencing datasets. This method is designed to effectively handle large patches of missing data, significantly improving imputation accuracy and reducing runtime for large-scale datasets.
Read full article
via arXiv — cs.LG
DB-3DME: From Dataset to Benchmark for Human-aligned Automatic 3D Mesh Evaluation
arXiv — cs.CV2 days ago
DB-3DME: From Dataset to Benchmark for Human-aligned Automatic 3D Mesh Evaluation
PositiveArtificial Intelligence
Recent advancements in 3D generation have led to the introduction of DB-3DME, a dataset and benchmark for evaluating 3D mesh quality, which includes 2,619 synthetic meshes rated by humans on geometry and prompt adherence. This initiative aims to address the limitations of existing evaluation methods in terms of cost, scalability, and alignment with human perception.
Read full article
via arXiv — cs.CV
Automatic Labelling for Low-Light Pedestrian Detection
arXiv — cs.CV2 days ago
Automatic Labelling for Low-Light Pedestrian Detection
PositiveArtificial Intelligence
A new research initiative has introduced an automated infrared-RGB pipeline aimed at enhancing pedestrian detection in low-light conditions, addressing a significant gap in available datasets and autolabelling processes. This method utilizes infrared detection to generate labels for RGB images, facilitating the training of object detection models like DETR, YOLO, and RCNN on the KAIST dataset.
Read full article
via arXiv — cs.CV
coding agents made repositories the security boundary
DEV Community2 days ago
coding agents made repositories the security boundary
PositiveArtificial Intelligence
GitHub has announced the general availability of security validation for third-party coding agents, including its own Copilot, Claude, and OpenAI Codex. This feature allows GitHub to run CodeQL, check dependencies against its Advisory Database, and scan for sensitive information when agents create code. If issues are detected, the agents attempt to rectify them, marking a significant step in ensuring code security within repositories.
Read full article
via DEV Community
MedicalRec: Medical recommender system for image classification without retraining
arXiv — cs.LG3 days ago
MedicalRec: Medical recommender system for image classification without retraining
PositiveArtificial Intelligence
A new study has introduced MedicalRec, a medical recommender system designed for image classification without the need for retraining. This system utilizes a dataset named MedicalRec-Bench, which includes over 5,000 records from 3,000 articles, focusing on various medical image classification tasks such as skin cancer and tumor classification.
Read full article
via arXiv — cs.LG
Benchmark Datasets for Lead-Lag Forecasting on Social Platforms
arXiv — cs.LG3 days ago
Benchmark Datasets for Lead-Lag Forecasting on Social Platforms
NeutralArtificial Intelligence
Recent research has introduced benchmark datasets for Lead-Lag Forecasting (LLF) on social platforms, specifically arXiv and GitHub. These datasets allow for the prediction of outcomes such as citations and forks based on early interactions like views and likes, addressing a gap in standardized datasets for LLF in the time-series community.
Read full article
via arXiv — cs.LG

Ready to build your own newsroom?

Subscribe to unlock a personalised feed, podcasts, newsletters, and notifications tailored to the topics you actually care about