Google’s open-source Genkit framework now ships a composable middleware system for TypeScript, Go, and Dart, with Python support on the way.
The system adds hooks at three layers of the tool execution loop:
<ul>
<li>Generate – conversation-level logic</li>
<li>Model – retries, fallbacks, and quota handling</li>
<li>Tool – approvals, sandboxing, and filesystem restrictions</li>
</ul>
These layers let you intercept what’s happening at each stage and apply your own logic before or after execution.
<h2>What comes out of the box</h2>
Genkit ships several pre-built middleware options:
<ul>
<li>Retries with exponential backoff – handles transient failures without manual wiring</li>
<li>Model fallbacks – automatically switches to an alternative provider when a quota is exceeded</li>
<li>Tool approval gates – adds human-in-the-loop checkpoints before a tool runs</li>
<li>Skill injection – pulls context from <code>SKILL.md</code> files into the agent’s workflow</li>
<li>Scoped filesystem access – restricts which paths an agent can read or write</li>
</ul>
Custom middleware follows a straightforward name-plus-factory contract. You can stack multiple middleware in explicit order, so the execution sequence is predictable. A content filter example in the docs shows how to enforce business rules deterministically in roughly 20 lines of Go.
<h2>Why this matters</h2>
Agentic applications have a reliability problem that prompt engineering alone can’t fix. When an agent is calling tools, hitting external APIs, and making decisions across multiple steps, you need infrastructure-level controls — not just better prompts. Middleware gives you a place to put that logic without tangling it into your core application code.
The Genkit Developer UI supports inspecting and tracing middleware execution, which helps when something goes wrong at 3am and you need to figure out which layer dropped the ball.
Full documentation is available in the Genkit docs.

Collections

Google's Genkit framework now includes a composable middleware system supporting TypeScript, Go, and Dart (Python coming soon). The system adds interception hooks at three layers: Generate (conversation-level), Model (retries, fallbacks, quota), and Tool (approvals, sandboxing, filesystem). Pre-built middleware includes exponential backoff retries, model fallbacks, human-in-the-loop tool approval gates, skill injection from SKILL.md files, and scoped filesystem access. Custom middleware follows a simple name-plus-factory pattern and can be stacked in explicit order. The Genkit Developer UI supports tracing middleware execution for debugging.

Genkit now has middleware for intercepting and customizing AI agent workflows