Operations | Monitoring | ITSM | DevOps | Cloud

Model routing promises to cut your AI agent spend by offloading routine tasks to cheaper models like Claude Haiku while reserving frontier models like Claude Sonnet for complex reasoning. In the right configuration, routing strategies can reduce inference costs by 40–85%. But if you implement routing incorrectly in a multi-turn agent, you can end up paying more than if you’d never routed at all. Here’s why and how to fix it.

Using JFrog CLI used to come with a catch. To run a build through it, you had to put it in front of the command. For example, jf npm install instead of npm install or jf mvn package instead of mvn package per the example below: npm install → jf npm install mvn package → jf mvn package Doing that once is easy enough.

AI agents are part of the modern development workflow. They write code, review pull requests, generate tests, call tools, interact with MCP servers, and help developers move faster. But behind every useful agent, there is something just as important as the model itself: the context that tells the agent how to behave. That context can include skills, prompts, instructions, hooks, commands, scripts, references, and MCP server definitions. In a small project, managing these primitives is pretty simple.

Today, we’re excited to announce that Boost is moving out of beta and into public preview. After months of building, breaking, and rebuilding inside JFrog’s own R&D organization, Boost is ready for the world. If you are currently running into token limits, unpredictable costs, or runaway usage from AI agents, Boost was built for you. It has already helped our teams reduce token spend while maintaining performance.

AI agents are changing how software gets built, but the infrastructure around them hasn’t caught up. Agents burn through tokens on noise. They take actions they shouldn’t. Context evaporates between releases. And most delivery pipelines were never designed for the pace and volume of agentic development.

Machine learning models are not like other software artifacts. A single fine-tuned LLM can weigh 70 GB. A model family may share 95% of its weights across dozens of variants. When hundreds of developers, training jobs, and GPU clusters all need the same model at the same time, the infrastructure underneath needs to be built for it.

Your developers install agent plugins every day: pulling from unmanaged GitHub repos, copying Cursor commands out of Slack, pointing Codex at a personal Git fork. Each of those is a new, uncontrolled distribution channel inside your software development lifecycle, and your platform team has zero visibility into any of it. A plugin is not a preference file. It is executable software, and right now it’s arriving on developer machines with no versioning, no provenance, and no audit trail.

You’re shipping a product with an AI-facing interface, or embedding AI-facing interfaces across your existing product line – skills your customers trigger, MCP servers their agent reaches for. Indie author or enterprise, your code runs in someone else’s agent runtime, against a model that updates every other day and a CLI that updates every other week. Cursor 2026.05.05-84a231c rolls out. Claude Code 2.1.132 lands the same week. OpenAI bumps gpt-5.5.

Since Anthropic announced Project Glasswing and the capabilities of Claude Mythos Preview, and OpenAI announced GPT-Cyber – my calendar has looked the same every day: Back-to-back calls with CISOs, AppSec leads, and security architects. And every call starts with the same question.

Developers building agentic AI on Google Cloud have powerful infrastructure at their fingertips: Gemini 3 for reasoning, Google’s Agent Development Kit (ADK) for orchestration, and a rapidly expanding ecosystem of Model Context Protocol (MCP) servers that connect agents to data and tools. So why are so many teams still waiting weeks to ship their first agent to production?