"We're building an agent!"
That's often where AI projects start getting expensive. Before you reach for an agent, ask four questions about the job that needs doing:
Metis · · 1 min read
1️⃣ Is it about understanding, writing or sorting information? → A language model is often enough.
2️⃣ Should the answers draw on your own knowledge? → Add search over your documents. Live data like order status? Call the system directly.
3️⃣ Does it need to touch your systems? → Decide on permissions and boundaries first. API or MCP comes later.
4️⃣ Does it need to choose its own path, use multiple tools and handle exceptions? → Now we're talking agent.
The hard part in production is rarely the model. It's deciding where it gets to make judgement calls – and where rules, permissions and approvals take over.
That's exactly how we built Metis: the agent takes the first pass, a human makes the call. Always.
Start with the job. Not the technology.