Agentic AI

A new arXiv paper challenges the assumption that compressing file reads saves money in multi-turn coding agents. Instrumenting a production gateway between agents like Claude Code and models like GPT-5, the authors decompose token bills into three levers. Controlled A/B tests reveal that tool-schema filtering, content compression, and history summarization save costs at fundamentally different rates.

• Aggressive compression preserves quality but doesn't guarantee lower bills in real sessions.
• Token savings vary significantly across filtering, compression, and summarization strategies.
• Instrumenting production gateways reveals the true cost attribution of each lever.
• Tool-schema filtering and content compression operate at different efficiency rates.

The AI-GRACE framework operationalizes agentic AI deployment by linking organizational governance and obligations directly to technical implementation details. It defines specific validation, control, and observation requirements to ensure use cases deliver intended outcomes while meeting compliance standards. The approach utilizes design science and situational method engineering to allow for contextual tailoring and reuse across different operational environments.

• Bridges the gap between high-level governance goals and low-level architectural controls for agentic systems.
• Focuses on defining what to validate, control, and observe rather than just model trustworthiness.
• Uses design science to create a reusable method for contextualizing AI risk and assurance.
• Aligns technical deployment capabilities with specific organizational obligations and use-case objectives.

Cloudflare has launched the Agent Development Lifecycle (ADLC) to streamline AI-driven engineering by supplanting conventional software development processes. This new framework targets inefficiencies in testing, deployment, and maintenance through automated software factories and dynamic orchestration. It also incorporates advanced observability and a dedicated security model designed specifically for autonomous agents.

• ADLC replaces traditional SDLC to address bottlenecks in testing and deployment for AI engineering.
• Automated software factories and dynamic orchestration are core components of the new lifecycle.
• A specialized security model is introduced to govern autonomous agent behavior.
• Advanced observability features are integrated to monitor agent performance and reliability.

This study evaluates efficient recurring evaluation methods for a production LLM analytics agent using 574 historical runs. Multidimensional 2PL adaptive testing delivered the highest score fidelity, achieving a 1.03 pp MAE with only 200 questions compared to a full run. Despite the superior accuracy of adaptive testing, the team deployed difficulty-stratified fixed subsets for operational stability.

• 2PL adaptive testing reduces benchmark load to 38.5% while maintaining high fidelity.
• Difficulty-stratified fixed subsets were chosen for deployment due to operational needs.
• Historical caching and random sampling offered lower fidelity than adaptive methods.
• Practical deployment experience highlights trade-offs between accuracy and stability.
• Recurring evaluation can be optimized without significant loss in score precision.