Why most AI agents never reach production
A model that predicts text isn't an agent that does a job. The gap between a demo and something real is a harness, and almost nobody builds one.
Every week there's a new agent demo that looks like magic. A few weeks later, almost none of them are running in production. I've watched this happen enough times, and shipped enough agents myself, to believe the failure is structural, not incidental.
A model is not an agent
A bare model call predicts the next token. That's it. It can't remember what happened yesterday, it can't touch your systems, and it can't recover when a step fails. A demo hides all of this because the demo is the happy path, run once, by the person who built it.
Real work is the unhappy path, run a thousand times, by people who didn't build it.
The missing piece is the harness
The thing that turns a model into an agent that survives contact with reality is a harness: the infrastructure around the model that lets it actually act.
- Skills: the concrete capabilities the job needs, specified and testable.
- Connectors: real access to real systems, with real auth.
- Memory: state that outlives a single call.
- Sandbox: somewhere to act without breaking things.
- Tools: well-typed actions with predictable failure modes.
- The loop: orchestration, retries, and evaluation baked in.
Miss any of these and you have a demo, not a product.
Build the harness, then the agent
The lesson I keep relearning: don't start from the model and hope the scaffolding appears. Start from the outcome you want, work backwards into a specification, and stand up the infrastructure first. The model is the easy part; the harness is the work.
That conviction is exactly why I'm building Kavela: describe the outcome, and get the harness a real job needs. Everything I learned shipping agents the hard way, turned into a platform.