The harness is all you need (mostly)

Vince Okay, so Burke Holland just wrote this thing called 'The Harness Is All You Need, Mostly,' and I'm reading it thinking… that's actually a pretty bold claim coming from someone at GitHub.

Ava Yeah.

Vince He's basically saying forget the prompt-engineering circus, forget the new skill everyone's talking about, forget the model of the week. The actual lever is just… understanding the harness. The core workflow. Using it well.

Ava I mean, that's not wrong. But it's also doing something interesting — he's basically saying the harness IS the infrastructure play. You learn it once, it works the same way in the CLI, the app, VS Code, JetBrains. That's the whole move.

Vince Right. And most teams haven't actually done that. They've got Copilot in five different places and they're learning it fresh every time they switch tools.

Ava Exactly. So the real productivity win isn't a new model or a clever prompt. It's depth. You get good at one tool, one harness, and you don't context-switch.

Vince Okay, walk me through the workflow he's laying out, because he gets pretty specific.

Ava So it's six steps. Pick a tool — doesn't matter which one, they all use the same harness. Then turn on YOLO mode, which is what he calls the allow-all setting. The agent executes without asking permission every time.

Vince Wait, right. And he's honest about the risk — you don't want to run that on your local machine with production data. You run it in a sandbox. Codespaces, dev containers, that kind of thing.

Ava Mm-hm. So step three is prototype. He gives this example of a date picker web component. Instead of diving into code, he says, 'Give me twenty mocks for a date picker. Put them all in an HTML file so I can compare.' And suddenly you see design variations you wouldn't have thought of.

Vince Because humans process visuals way faster than dense text. That's not a new insight, but it's… it's a reminder that the AI can generate mocks and diagrams instantly. Why not use that?

Ava Exactly. Then he switches to plan mode. Still the same session, same model, so prompt caching is still working. And he asks the model to build out the plan — all the edge cases, all the questions he'd need to answer himself anyway.

Vince Can the start and end date be the same? Should users be able to clear the date? Is manual entry allowed? That stuff.

Ava Right. And here's the key part — he's not accepting every suggestion. He's engaging deeply. If the model asks about non-contiguous dates and he's not sure what it means, he asks for clarification. The planning step is where his expertise actually matters.

Vince Stop—

Ava —okay?

Vince That's the part that actually interests me. Because he's saying the harness is all you need, but what he's really describing is a workflow where the HUMAN is still driving. The planning step is not automation. It's using automation to make the human smarter.

Ava Yeah, that's fair. The harness doesn't remove judgment. It surfaces the questions that require judgment.

Vince Okay, so then step five is Autopilot. Once the plan is done, the model starts implementing. And here's where the infrastructure gets interesting — the app is automatically routing to different subagents depending on what it needs to do.

Ava Also, the 'no weird prompts' thing is marketing. Good prompts still matter. The planning step is where you're still writing prompts. They're just more methodical instead of hoping for magic.

Vince But here's what I think lands. Most teams ARE trying new models, new skills, new tools every week. And meanwhile, nobody's actually gotten good at the one tool they have. This is a 'go deep instead of broad' argument.

Ava And it's not a new argument. But it's useful to hear it from someone shipping this stuff at scale. The harness is the thing that doesn't change. Everything else is churn.

Vince Last step is human review and iteration. The model generates something. It's probably not quite right. You iterate until it is. And he's honest — the date picker the model gave him had color contrast issues, inconsistent animations, design choices that screamed 'made by AI.'

Ava That's the part where taste lives. The model can build the structure, but you're still the one who decides whether it's good.

Vince So the real take is… the harness is infrastructure that lets you spend your energy on the parts that matter — discovery, planning, judgment, iteration. Not on approving every single action the agent takes.

Ava Right. And the infrastructure is boring. It's routing, prompt caching, subagent selection. None of that's flashy. But it's the difference between the workflow being useful and the workflow being exhausting.

Vince I think what gets me is that this is the flip side of the agent hype. Everyone's trying to build smarter agents. Holland's saying, no, you need better harnesses. Better infrastructure around the agent. Make the agent autonomous enough that you're not drowning in approvals, but not so blind that you lose visibility.

Ava Yeah. And the harness does that by design — it forces you to think through requirements early, plan methodically, and then let the agent do the work. If you skip any of those steps, the whole thing breaks.

Vince The question I'd ask is whether this generalizes. Like, does this workflow work for debugging a production outage? Or refactoring a legacy system?

Ava Probably not the same way. This is optimized for building new things. For brownfield work, you need more context, more constraints, more human judgment upfront. The harness can help, but it's not going to autopilot you through a minefield.

Vince Fair. But the underlying insight — learn the tool deeply, use the same model for the duration of the task, let the infrastructure do its job — that probably applies everywhere.

Ava Yeah, that's portable. And it's the opposite of the 'try everything' energy that's been dominating the discourse.

Vince This is the kind of piece that's going to be useful for teams that are already bought in on agents but feel like they're drowning in process. Someone's going to read this and go, oh… we don't need a new tool. We need to actually learn this one.

Ava And they're going to save real money on tokens just by not model-hopping mid-session. The prompt caching thing is real.

Vince Yeah, okay. This landed different than I expected. I thought it was going to be 'GitHub is promoting their own product,' but actually it's… it's pretty honest about the tradeoffs and where the workflow breaks.

Ava That's because it's written by someone who actually ships. Not marketing. Just… here's what I've learned.