My relationship with LLMs has moved through three phases, and each one changed how I think the leverage actually works.
Phase 1 — The proxy
It started with the AI Proxy Platform and MecutinAI: one /v1 endpoint across
many providers. The insight here was plumbing — collapse provider sprawl into
one contract. Useful, but it's infrastructure. The model is still just a
function you call once.
Phase 2 — Planning as a product
Then VibeEngine: turn a raw idea into an AI-ready blueprint (PRD, DESIGN, TECH_STACK, AGENTS.md) in a few minutes via a 6-step wizard. The shift here was structured context. A model with a messy prompt produces messy output; a model handed a well-shaped context produces sharp output. The product wasn't the model — it was the scaffolding that produced the context the model needed.
That reframed things for me: most "prompt engineering" is really context engineering. The prompt is the last mile; the structure feeding it does the heavy lifting.
Phase 3 — Agentic workflows
This is where I'm betting now. The difference between a single LLM call and an agentic workflow is that an agent loops — it acts, observes, and acts again with tools. That's powerful, but it's also where things get fragile:
- Small, focused agents beat mega-agents. One agent that retrieves, one that drafts, one that critiques. Compose them. A single agent trying to do everything drifts.
- Evals guardrail everything. An agent without an eval loop is a vibe-check. The agents that actually ship run through cheap, automated checks before returning — and fall back gracefully when they fail.
- Tools are the moat, not the model. Swappable models are a commodity; the
tool surface (what the agent can actually do in your system) is the
differentiator. The MecutinAI gateway taught me this by accident — the value
was the contract + the tools around it, not which LLM sat behind
/v1.
Where I'm betting
- Context engineering over prompt tricks. Shape the input, don't beg the model.
- Composable small agents + eval loops, not one giant autonomous one.
- Tool-rich systems where the model is the cheapest swappable part.
- Human-in-the-loop on the risky transitions — agents that can ask, not just guess.
The trajectory is clear to me: the model layer is flattening into infrastructure, and the leverage is moving up — into how you structure context, compose agents, and wrap them with tools and evals. That's the layer I'm building toward.