Agentic coding, using AI tools like Claude Code, Cursor, or Codex to write software, is an emerging and exciting area of tech. If you’re in any way involved with building software, it’s probably one of the most important conversations you’re having.
There’s no lack of momentum or enthusiasm for these new ideas. There’s also a swell of failures and financial pushback.
While there are many ways to make mistakes during this paradigm shift, this swell of problems mostly boils down to a single mistake made over and over again. In this post, we’ll talk about how to avoid it yourself.
The Current State: Agentic Coding
Where we sit in the summer of 2026, agentic coding is both exciting and fraught. On the one hand, you’ve got forward-thinking software engineers challenging decades-long assumptions in real time and in public on platforms like X. They are producing anecdotal evidence, even if not exactly replicable for you today, that agent-led software engineering can be a highly effective (on a 10-100x scale!) new way to build software.
On the other hand, your social feeds are probably also littered with stories like Uber burning through its entire AI budget in months. You see pronouncements for the oncoming AI crash, because businesses are just going to find it too expensive. And, frankly, you very likely have your own finance department starting to ask questions!
Most companies dove into this without much thought. In their defense, that’s often the case when a new technology rapidly emerges. But, in this case, AI’s usage-based pricing, revved by a compute supply constraint, means the cost of winging it is virtually unbounded, and it’s starting to spook CFOs.
That means we have massive excitement for the biggest reset for software engineering in decades, and it’s coupled with massive hesitation in the boardroom.
The Mistake: Implementation Without Intention
There’s one mistake that shows up in almost every implementation. Entering the initiative with little understanding of your intended outcome is the biggest driver of whether an effort to incorporate agentic AI will be scrapped.
You may think you know what you want out of the initiative. I commonly hear things like, “We want our engineering team to go faster!” or, “We want to improve quality.” But those are often justifications teams use to cover up the real desire: “This technology is just really cool.”
The problem with all of these reasons, even the ones that are directionally correct, is that for most teams they are not observable or measurable outcomes. What does going faster mean? How is higher quality measured? You and your team may all generally understand what these objectives mean, but your CFO probably does not.
That CFO will come into your office and ask, “We spent $50k on Claude this month? For what!?” And the problem that is hampering success (even killing AI rollouts) is that almost no one has a real answer.
The Fix: Begin With The End In Mind
This mistake was made at the beginning, before you signed up for Claude or Codex, and before you consumed a single token.
The feedback loop is slow. It’ll take a quarter or two before your CFO starts asking questions. That delay means it’s hard to even see that you’ve made the mistake.
The most important thing you can do, as early in the process of rolling out agentic coding tools as possible, is to define one to three goals in measurable (or at least observable) terms. Then measure your baseline.
You want your engineers to go faster? What does faster mean? Do you have a mature enough Agile process for story point velocity to represent that? Or do you mean feature cycle time? Who all is included in going faster? Is it just developers, or do you mean the full product management-to-release SDLC? What is the number you are going to measure, or how will you observe if the outcome was achieved?
Ideally, this is a number. Even better, it’s a number you can translate into tangible dollars for the business. But that isn’t always feasible, and it doesn’t mean the benefit isn’t there. That’s why objectively observable outcomes can also work.
Think of it this way… When your CFO walks into your office with a hefty invoice from your agentic coding tool, and asks you what the company got out of this bill, what do you want to lay out for them? What story do you want to be able to tell? That story should define these observable or measurable outcomes.
And, the most important time to have this conversation is before implementing an agentic coding practice. It’s ok if you get it wrong and have to adjust. You probably will. But this conversation as a reaction to that cost challenge makes it really hard to establish an unbiased success framework. You’ll be too tempted to back into numbers that feign justification for the money your team already spent.
Define the outcome you want at the beginning!
The Bottom Line: Define It Before You Deploy It
Agentic coding is easily one of the most exciting shifts in software development today, but momentum alone won’t guarantee success. As the initial hype gives way to financial scrutiny, the line between a scrapped AI initiative and a successful transformation comes down to clarity.
If you want to avoid the single biggest mistake that tanks these projects – launching into agentic AI just because it feels “cool” or because you want a vague boost in “speed” – you have to know exactly what you are aiming for.
Before you roll out tools like Claude Code or Cursor across your engineering organization, take a step back and define one to three specific, measurable goals. Capture your baseline metrics today so you can actually prove the impact tomorrow. By shifting your approach from chasing trends to tracking concrete outcomes, you ensure that your investment in agentic coding delivers real, undeniable value.