Pair-programming Agentic Financial Applications with AI Agents

Here’s the revised version:

Pair-programming Financial Applications with AI Agents

The IGV ETF, the benchmark for software sector stocks, has dropped more than 20% since the start of the year.

Enterprise SaaS names that looked untouchable eighteen months ago are being re-rated in real time. “SaaSpocalypse, Brutal Wipeout, 2026 SaaS Crash, AI-driven software stock crash” were among the headlines over the last weeks.

The market is asking a question it hasn’t seriously asked before:

If AI can write software, what exactly are we paying for?

In my opinion, the answer is more nuanced than the headlines suggest. Data / Network effects, switching costs, compliance requirements, and distribution moats don’t evaporate overnight.

But the act of writing software? That value may already be approaching zero.

The way financial software gets built is changing fast. And I don’t mean incrementally.

Over the past month, I ran a personal experiment: build a non-trivial, production-grade agentic investment management system from scratch, using two of the leading AI coding assistants, Codex-5.2 and Claude Sonnet 4.6, as my pair-programmers.

The system I set out to build wasn’t planned to be a weekend prototype.

It includes a multi-agent reasoning loop, a three-layer persistent memory architecture (episodic, semantic, and procedural), a real-time sensor layer pulling from sources like SEC EDGAR, Senate disclosures, and earnings calendars, and fourteen analytical playbooks covering everything from fundamental underwriting to exit discipline.

In other words, the kind of system that, eighteen months ago, would have taken a small team several sprints to architect and ship.

What I found surprised me.

The gap between the two coding assistants was significant in code quality, architectural coherence, and cost. One felt like a junior developer: fast, eager, but generating messy output that required constant auditing. The other felt like a senior: more deliberate, cleaner code, fewer rollbacks. The cost difference was 11x.

But beyond the model comparison, the deeper observation is this: the translation layer between idea and working system is collapsing.

My conclusion is that the bottleneck is no longer technical execution. It’s the clarity of your vision. Or in terms of the SDLC: while development time is reducing, requirement time and testing time are increasing.

For financial services professionals, whether in wealth management, asset management, or fintech, this will have real implications. Finance is Math and Math is code.

The Anthropic CoWork wealth management plugin (which triggered this whole experiment) is a signal of where enterprise tooling is heading: AI embedded directly into existing workflows, not bolted on as a separate chatbot.

The incumbents who move aggressively on AI-augmented development will ship faster, run more experiments, and improve margins. Those who don’t will face erosion, not overnight, but steadily.

I wrote up the full architecture, the agent design, the sensor layer, and the detailed comparison between the two coding tools over on my publication, Encyclopedia Autonomica.

If you’re building in financial services or thinking about where AI development tooling is heading, I think you’ll find it worth the read.

👉 Read the full post on Encyclopedia Autonomica

Curious whether others in the FinExtra community are experimenting with AI-assisted development for financial applications and where you’re seeing the sharpest productivity gains (or disappointments).