The Agentic Coding Moment – Do You Know Who’s Ready?

Editor’s note: Welcome to a new series for engineering and talent leaders in financial services. You may be seeing this because you’ve connected with someone on our team or are part of our professional network.

In the last few weeks, I’ve seen a sharp uptick in inquiries from financial services executives about AI workforce strategies. As headlines about agentic software development rattled software stock prices, executives are looking at their organizations and asking: “How do I know if my engineers are good and will they succeed in this new world?”

These are fair questions, but the answers aren’t simple.

In this article, I’ll share:

• Why those questions are surfacing now
• The two dimensions to AI talent: AI-enabled vs AI-specific
• The two things every financial services technology leader can do now

The AI “lightbulb” moments

The future is already here; it’s just not evenly distributed. – William Gibson

Something has shifted in the last few weeks. Not only in the technology itself, but in how people are using it and who is recognizing its potential.

The foundational models are continuing to improve exponentially. But a lot of that can sound like noise after three years of relentless AI hype.

What’s different is that some of the most credible, battle-tested voices in software engineering are converging on the same conclusion, independently, and with increasing conviction: human-written code is dropping to zero, with some claiming the software delivery lifecycle is dead.

Here are some of the chorus of voices I’ve seen circulating heavily, across technology leaders and financial services tech leaders alike:

• On New Year’s Day, Steve Yegge published “Welcome to Gas Town.” The 40-year engineering veteran (ex-Amazon, ex-Google) laid out eight stages of AI-assisted coding and described a future where developers manage fleets of AI agents the way factory operators manage production lines. It was provocative. It was also heavily caveated. Just six weeks later, Steve’s Pragmatic Engineer interview dropped every caveat. 
• Matt Shumer, CEO of OthersideAI, published “Something Big Is Happening,” a nearly 5,000-word warning that went viral, drawing over 80 million views in a matter of days. Shumer believes he is no longer needed for the core technical work of his own company. He sees AI leaping into something resembling judgment and taste. According to Shumer, AI is now a general substitute for cognitive work.
• Spotify’s co-CEO, Gustav Söderström, revealed that the company’s best developers haven’t written a single line of code since December. Their internal AI system handles implementation end-to-end. Engineers describe what they want through Slack, AI builds it, and they review and merge. 
• Meanwhile, Goldman Sachs partnered with Anthropic to build AI agents for trade accounting and client onboarding. Goldman’s CIO admitted he was surprised at how well Claude handled complex, rules-based work far beyond coding.

While the tech vanguard talking about this may be somewhat expected, the biggest shift I’m seeing is in the corner offices. Executives, both CEOs and technology leaders, are getting hands-on with AI-assisted development and experiencing it themselves. Whether they’re building hobby apps or ramping their own code commits, the effect is the same: the lightbulbs are turning on.

AI Talent Needs: Two Dimensions

That urgency is showing up in my inbox. The pace at which financial services technology leaders are reaching out has meaningfully accelerated. They’re looking for guidance, looking for a framework, looking for answers. And when I dig into what they’re actually asking about, the questions are clustering around two distinct talent dimensions.

1. How do I assess and develop my AI-enabled engineers?

This is where the majority of existing workforces reside, both full-time and contracts. Copilots have become the norm now — Claude Code, Devin, Cursor, etc. But the definition of AI-assisted is rapidly expanding to include increasingly sophisticated orchestration of AI agents and subagents.

Yegge’s eight-stage framework is the best map I’ve seen for understanding where an organization stands. At the bottom (stages 1, 2, and 3), engineers are barely using AI, maybe accepting some code completions or occasionally asking a chatbot a question. By stages 4 and 5, they’re working primarily through AI agents, using tools like Claude Code or Cursor as their primary interface. But the real leap happens at stages 6 through 8. The engineer is no longer working with a single AI assistant. They’re orchestrating teams of agents working in parallel across a codebase: setting direction, reviewing output, resolving conflicts, maintaining architectural coherence.

If that sounds less like coding and more like management, that’s because it is. As I posited recently on LinkedIn, the maker/manager distinction Paul Graham made famous is collapsing. The skills that matter most at stages 7 and 8 are the ones we’ve traditionally reserved for senior engineering leaders: communication, prioritization, architectural judgment, the ability to articulate what to build with enough clarity that others — human or AI — can execute on it. Deep technical knowledge hasn’t become less important. It’s become important differently: the engineer who can orchestrate agents but doesn’t understand system design or failure modes will produce confident, fast, wrong output at scale.

At the end of the day, AI is becoming an integral part of how software engineering teams are working. There is a spectrum of depth for what this looks like, and a range of management-like skills becoming more important to all AI-enabled engineers.

Separately, I’m also increasingly getting questions about AI-specific engineering roles.

2. How do I build out AI-specific engineering capabilities?

This is a different challenge entirely. These range from the deep experts who can fine-tune models, build the platforms that govern how thousands of colleagues safely use AI, and construct the guardrail infrastructure that regulated enterprises need. These roles may be few, but their leverage is extraordinary.

They’re also more likely to be net-new skills you need to bring in-house, which means few people internally know how to evaluate them. The quality bar is very high and the competition for this talent is fierce. JPMorgan, for instance, seems to have broken from its own compensation norms to pay AI engineers at a premium over peers. It’s a clear signal that they understand there is scarcity in the supply for this talent. Initiatives like OpenAI’s Frontier Alliances underscore the shortage in supply and the role strategic partners will play to fill the gaps.

Win the Middle

Some of your workforce is already embracing AI for technical work. They’re pushing to get the tools, building hobby apps on weekends, and learning fast.

Another camp is resisting hard. They’re naysaying — sometimes for good reason, since the current way can occasionally be faster than the new way — and they’ll be the last to change.

Then there’s the middle camp. Depending on your organization, this could be eighty percent of your technical workforce. They’re willing to change, but they need to know it’s safe. They need to see it work — in production, at your company, by people they respect.

This goes beyond introducing AI tools. If leadership can’t catalyze a mindset shift, inertia compounds. The leading camp gets frustrated and leaves. The culture of experimentation dies. The gap widens. Once you’re on the wrong side of this momentum, catching up gets exponentially harder.

We’re living this transformation at Karat alongside everyone else. But we’re seeing patterns that lead to success. Our leading camp is leading by example — we’re intentionally bringing in engineers who operate at the edge and selecting a few to lean in hard on AI-native development. Those engineers prove to the rest of the organization what can get done in a production environment, in 1/10th the time, at similar or better quality. For the middle camp, the mindset shift has to be experienced to be believed. You can’t memo your way into it.

What to Do Now

Step 1: Know thyself, know thy people

You can’t build a map to get where you’re going if you don’t know where you are. 

Mis-measuring talent during a structural shift risks misallocating capital and making the wrong AI bets.

Engineering leaders already struggle to know the quality of their workforce. Almost none have measured it against the capabilities that matter now. Most assessment frameworks — interview rubrics, performance reviews, promotion criteria — are still calibrated to a time when writing correct code was the job.

That world is gone. If your rubrics don’t assess architectural reasoning, tradeoff analysis, and the higher-order judgment needed to orchestrate a fleet of agents, you’re measuring the wrong thing entirely. We need updated technical rubrics to assess AI-assisted technical talent.

Step 2: Let your leaders prove it

Once you know where your people stand, use that information. Find the engineers already operating at stages 6, 7, or 8 and put them in front of the skeptics. Give them real problems, in production, with visibility. The middle camp doesn’t need a memo — they need a colleague they trust to show them it’s real.

Lighting the path forward

The lightbulbs are turning on. The executives who get there first won’t win because they moved fast — they’ll win because they knew where their people actually stood, and they put their best proof points in front of the skeptics. Measure who’s ready. Let them lead. The middle camp will follow.

Jeff Spector is the Co-Founder and President of Karat, the trusted standard for measuring talent quality. Karat helps engineering leaders at companies like PayPal, Atlassian, and Citi make their most important talent decisions through human-led, AI-native talent evaluation.