I do. Harness engineering is more expensive, but let me walk you through my thought process.
When you put an AI agent to work on something real, there's a whole system underneath that decides what it's allowed to touch, what information it actually has access to, and how you confirm the work is done before calling it done.
That system is the harness.
Building one takes time, and yes, it costs more than just letting the agent run.
But, and let me hold your hand when I say this, I strongly believe that an agent without a harness isn't working cheaper, only unsupervised.
I recently read a test where two agents were given the same task.
The first one ran solo, no real structure around it. Finished in 20 minutes and cost about $9. It looked done, but hello, hello, there was a bug in it that only showed up if someone actually read through the code * _ *
The second had a proper harness. A planner, a builder, a checker. It took 6 hours and cost about $200. More than 20 times the cost, but it actually worked.
So, the first run didn't save money, it just made the mistakes harder to see.
A bug that makes it to production, a rollback, the hour someone spends reconstructing what the agent did and why. The bill shows up eventually.
Harness engineering is really just the same instinct you already apply to shipping software, except the thing you're managing is an agent.
The elements are roughly what you'd expect: what the agent can do without asking, what context it has access to, how it holds state across steps, and how you actually verify the output before anything gets marked complete.
We have a recent video on this that even walks through a basic permissions example in Claude Code if you want to see what the smallest version of a harness looks like in practice.
That’s my take on the whole harness engineering debacle, what’s yours? Am I just burning budget?
Top comments (0)