Build systems where you can plug in any LLM and test fast.
Your work is testing prompts and ideas—not fighting infrastructure.
I've shipped AI agents to production for marketing tech over the past few years. What works:
- ✅ Model switching takes one line of code
- ✅ Prompt editing is instant
- ✅ Zero adapter layers or vendor lock-in
- ✅ Testing happens in minutes, not hours
The Principles
Design for model flexibility
Flagship models drop constantly. Your system should plug in any LLM without rewiring. Switch providers to test. Swap models to compare. Infrastructure stays invisible.
Optimize for prompt iteration
Your velocity comes from rapid prompt testing. Combine evals with real-world usage. Iterate fast. Test rigorously.
Keep everything else out of the way
No abstraction layers. No framework overhead. Direct LLM integration. You should spend your time on prompts and ideas, not plumbing.
Stack I Use
Model Routing & Testing
- AI Gateway — Switch providers without touching application code. Built for Vercel Stack.
- OpenRouter — Access hundreds of LLMs through a single API. Great for rapid comparison. I use this with n8n.
Workflow & Orchestration
- AI SDK — Core workflow orchestration and agentic systems. Clean abstractions without sacrificing control.
- Workflow — Long-running, durable workflows on Vercel.
- Trigger.dev — Background jobs with retries, observability, and task chains.
- n8n — Visual automation that non-technical team members can modify. Great for prototyping before going to production. Self-host it on Railway to scale to millions of actions per month.
Agent Tools and Frameworks
- MCP — Provides tools to agents. Not perfect, but best we've got.
- Mastra — Production-grade agent framework. Handles coordination so you focus on logic.
This stack ships agents that work.