My portfolio headline says I build "agents, harnesses and loops." People ask about the agents. They should ask about the harnesses, because that's where production AI work actually lives.

The model is the smallest part of the system

Working on social listening pipelines at Yap Market taught me this early. The LLM scoring step, the "AI part", was a fraction of the system. Around it sat the real work: fetching tweets reliably, matching them against campaign context, maintaining context windows so the model saw the right information, syncing results in real time, and keeping all of it running when upstream APIs misbehaved. If you only evaluate the model, you're evaluating the easy part.

What a harness actually does

The harness is the deterministic layer around the model. The model proposes; the harness disposes. In practice that means:

• Validation: checking that what the model wants to do is well-formed before it touches anything real
• Budgets: caps on calls, tokens, money, and time, enforced outside the model where it can't talk its way past them
• Recovery paths: when a step fails, the harness decides whether to retry, reroute, or stop, instead of letting the model improvise
• Observability: logging every proposed and executed action, because you cannot debug what you didn't record

None of this is glamorous. All of it is the difference between a demo and a product.

Loops need stop conditions before they need intelligence

Every agent loop I've shipped got better when I made the stop condition explicit first: a maximum number of iterations, a budget ceiling, a concrete definition of done. An agent loop without a deterministic stop condition isn't autonomous, it's unsupervised. Those are different things.

Voice makes all of this harder

Voice agents compress every harness problem into real time. There's no retry button in a live conversation: latency is a feature requirement, interruptions are control flow, and a hallucinated answer is spoken aloud before you can filter it. Building voice onboarding for a fintech taught me to design the guardrails first and the personality second.

Where I've landed

When I evaluate any agent system now, mine or anyone else's, I ask three questions. What does the harness refuse to let the model do? What stops the loop? And what happens when step three fails at 2 a.m.? The answers predict production success better than any benchmark.