Agentic AI, beyond the hype
Everyone is talking about AI agents. We're building them. Here's what we've learned about what works, what doesn't, and what's actually ready for production.
Draft placeholder. Full article pending author sign-off.
Agents are having a moment. The vendor pitches make it sound as if it's solved. We've shipped enough of them to know it isn't, and to have a view on which kinds are ready for production work and which aren't.
What actually works today
Document-heavy workflows: intake, extraction, classification, routing. Anything where a human currently reads a stack of PDFs and makes a repeatable call. These agents ship reliably now. The Claude Agent SDK plus a decent tool set gets you most of the way.
What nearly works
Multi-step processes that span several systems. The LLM reasoning is usually fine; the hard part is giving the agent stable tool access. MCP has changed this significantly over the past year. We expect this category to ship cleanly in 2026.
What doesn't work yet
Open-ended planning tasks with long horizons and expensive failure modes. Any role where you'd hesitate to give a junior hire full autonomy, you should hesitate to give an agent too.