I remember carelessly writing it down in my notes. That was the entire extent of my response. I didn’t think much of it, and now agents are what everyone is talking about and I’m late to the party; all because I heard something and mindlessly wrote it down with no action plan.
I’d been around AI for a while by then. Computer vision projects, machine learning coursework, using the tools like everyone else. For some reason I never went into agents the way I should have.
So this summer, as an intern at one of Deloitte’s technology arms, I built one. A local, permissioned autonomous agent, from nothing, over about two months.
That is the honest starting point. I had heard the word, but what did it actually mean? I had no idea.
I had heard of the LangGraph family without really knowing what it was. I knew containers existed as a concept, via Docker, but had never used one. I understood the Model Context Protocol well enough to explain it and had never applied it to anything real.
A harness, as I now finally understand, is everything around the model. The model is one component and arguably the least interesting one. The harness is what decides which tool gets called and what happens when that tool fails. It also determines what the agent is allowed to do without asking, where generated code is permitted to run, and what happens when the model produces something confident and wrong.
All of this was news to me because before this internship, I’d just look at benchmarks. This year alone, people went crazy when Gemini 3 came out and within a few days, OpenAI had clapped back with GPT-5.2. Then we saw Anthropic take the lead for months with reports of them not giving the average Joe the best models, but they quickly changed their mind when the Chinese open-source models like Kimi K3 and Qwen 3.8 (which can run locally) emerged. All of this does matter and I won’t stop keeping up with these updates. But for someone like me, I’m not planning on creating my own LLM so I’ll just use the best frontier model that’s cost-efficient at the same time. Long story short, I was focused on the wrong part of the AI stack.
The single idea that shaped everything else in my agent:
The LLM writes text. Deterministic code makes the consequential decisions.
The model drafts the plan. It does not pick the tool. Tool selection runs through code I can read, test, and reproduce.
The thing is, a model is probabilistic. When it picks the wrong tool, every node downstream trusts that pick and builds on it, and what comes out the other end is a confident hallucination. Unlike runnable code where I can get an error message, I get absolute confidence and very unsafe behavior from the AI.
The worst version of this in my system: someone says “send the email” and the router quietly selects “search emails.” It looks like it worked. Nothing throws. The email was never sent.
So I built a deterministic guard that classifies write intent separately from the model, and vetoes a read-tool selection when the user clearly asked to write. Not because it’s elegant. Because that failure is silent, and silent failures are the ones that hurt.
My tool routing started as a large regex layer which consisted of roughly 1,500 lines turning phrasings into tool calls. It worked… kind of. I thought the project would end there and confidently told my boss I was basically done, but little did I know that this was just the beginning.
My manager told me regex was a bandaid and that it wasn’t scalable. So it was back to the drawing board.
He wasn’t wrong, and my first instinct was to argue. What I did instead was go find out what frontier labs actually do: trained function-calling, structured outputs, and evaluation harnesses and then build that version on my local model, keep the regex behind a flag so the whole thing was reversible, and measure which one was better.
That reframing is the most useful thing I took from the internship. The question stopped being “is regex bad” and became “bad compared to what, on what data, by how much.”
I wrote 488 prompts and split them: 300 to tune on, 188 sealed in a drawer. Thresholds were swept on the training split only. The hold-out got opened once, at the very end.
My regex router had been scoring near 100%.
On the sealed set it scored 51.1%.
That gap is the entire lesson. The old number wasn’t a lie, it was train accuracy on its own examples — I had tuned the regex against the same phrasings I was testing it with, and then believed the result. It’s the oldest mistake in machine learning and I walked directly into it while thinking I was being careful.
The cascade router I replaced it with used a local model first, a second model behind it, and regex only as a tiebreak. It scored 89.9% on the same sealed set. More importantly it got read-versus-write right 100% of the time, against 19 dangerous misroutes from the regex.
Worth being precise about what that does and doesn’t mean: 89.9% is against my set of roughly 45 tools, using a 7B model running locally on my machine. A frontier model would score higher. That’s rather the point my manager was trying to put forward the whole time. The interesting result isn’t that it’s state of the art, it’s that a small local model plus a carefully built harness was good enough to beat the thing I’d hand-tuned for weeks.
The harness did more work than the model did.
Here is the thing I keep circling.
The same way AI is a loop of processes, the use of AI is a circle for a consumer like me as well. I write code with AI. I ask questions about my code to AI. I brainstorm with AI. And then I sat down to build an agent and a harness whose entire purpose is deciding when to trust a model’s output… using a model to help me build it.
Am I validating my work with AI, or walking into a loop of bias I should be avoiding? Is this inherently wrong? I don’t think so. But it does dial in on an important matter and it’s that my judgment is more important than ever. I can now build much more impressive things and much faster due to AI, but with that comes responsibility to make the right decision to scale a product/solution safely and efficiently. And yes, AI can help with that too but if you’re blindly following it, you’re the one to blame in the end. And I think I’ve seen enough stories about companies in the consulting industry having multi-million dollar cases brought against them because of just that.
When I asked AI whether my routing approach was sound, I got agreement. Useful agreement, even. But agreement is what that interaction produces, and I can’t tell the difference between “this is correct” and “this is plausible” from inside the conversation.
In my project, I saw this when the sealed hold-out told me I was wrong. 51.1% is not an opinion and it did not care how confident I was. My manager saying “regex is a bandaid” was outside the loop too and must be considered.
That’s the rule I’ve landed on. Use AI to build, absolutely. But every claim that matters needs a check that can independently return “no.” A test that fails. A hold-out you only open once. A person who’ll tell you the thing you don’t want to hear. Without at least one of those, you’re not validating anything, you’re just generating more confidence.
I’m in no way an expert in any of this. If anything, building the thing showed me how far I am from going from a novice to a master. I now know what a harness is, which mostly means I know how much of it I haven’t built yet.
And as a final note, the biggest lesson: AI is only as good as you are. Coding agents are insanely good now, but after trying to build my own harness and seeing how stupid the LLM can be was a wake-up call that if you just depend on the model’s intelligence and whatever frontier model is going to be released next Tuesday, you’re just falling behind.
The agent is open source: langgraph-code-agent — the cascade router, the evaluation harness, and the methodology write-up with the full per-service breakdown are all in there.
]]>