The New Craft: Building the Harness

Michael Laccetti, Senior Director of Engineering at Gorgias, on how the engineering job moves up a layer when AI writes the code, and the two costs that come with it: one paid in people, one paid in tokens. An audio version of Michael Laccetti's interview is available on Spotify.

Three Takeaways

• Building a harness is now the craft. When agents write the code, the engineering work becomes building and maintaining the harness.

• The harness has a people cost. Redefining the job from writing code to writing specs, building the harness, and reviewing an agent's work can be seen as a loss for many engineers whose craft was writing code.

• An even bigger token bill could be on the way, but there are solutions to mitigate cost that every org can take.

Introduction

Over 25 years, Michael Laccetti has gone from IC to principal engineer to running a whole org. He is now Senior Director of Engineering at Gorgias.

Laccetti argues that when you delegate writing code to an agent, guardrails and harnesses matter more than ever and should now be the core part of engineering work.

The Harness Has Become the Craft

Laccetti's team had one structural advantage: they were building greenfield. Pointing AI at a brownfield codebase, he notes, carries real overhead because the model has to absorb the existing architecture and its interactions before it can help. With less legacy and tech debt to account for, his team could primarily focus on integrating coding agents into their everyday systems.

They started where most teams do, handing off more manual work where they could. Unit tests went first, but they quickly learned an important lesson: if the model doesn't understand the code, or the business problem behind it, the tests aren't worth much. Prompts and context, Laccetti found, mattered as much as they ever had.

The first structured fix was skills, small written instructions scoped to a single unit of work. Their limit, in his telling, is that they are only words, and words "can be ignored or misunderstood or misinterpreted." A model that has settled on its own reading of what you meant will follow it confidently in the wrong direction.

So the team built plugins with deterministic hooks: enforced schemas, fixed approaches, and judges that check the output at each step and escalate when something looks off. The aim, Laccetti says, isn't a smarter model but one with fewer places to go wrong. It's a direct answer to a failure mode he keeps hitting: "when they're convinced of something, they'll just keep running with it." The deterministic checks are what stop a confident model from running off a cliff.

In short, Laccetti's team was building the harness: deterministic software wrapped around a probabilistic model.

The Costs of Building a Harness

But there are costs to building a harness, and not all of it is technical.

Now that agents are writing code, most of the work has transitioned into specs, review, and harness work. For some engineers that’s a fine trade. For others it's seen as a loss ("I didn't come here to write documentation" is a reaction that Laccetti has heard). Laccetti's approach is to find tasks and experiments that give each engineer room to grow and match their interests to the best of his ability.

The other cost is literal. Running agents at this volume – with multiple reviewers and agents at most if not every step – uses tokens, and the bills only grow as contracts move to usage-based pricing. Most teams are still guessing at budgets, and Laccetti's view is that there is no universal answer.

His plan for when the meter starts running is an LLM gateway, sending work that he can to a cheaper model and saving frontier models for higher-level work. Many teams, he believes, have already begun adopting this exact solution.

Engineering in a New Form

Laccetti's own read is that AI simply exposed system-wide issues that always existed among engineering teams. Fixed capacity, where to invest that capacity, and how to validate the result were all long-standing questions well before agents were introduced. 

What's changed is that the same questions now show up on day one. The guardrails, safety nets, and platform investment that used to belong to companies with thousands of engineers are becoming table stakes for a team of ten. With that perspective, the harness isn't something you bolt on once you've scaled. It's the foundation you start from.

For more episodes of Stack Trace, subscribe to Span's YouTube channel or to the Stack Trace podcast on Spotify.