🤖 AI-Assisted Product Development, in Practice

February 06, 2026

By Ted Steinmann

The Starting Point

This started with a familiar frustration—returning from a trip and facing yet another spreadsheet to submit expenses to an accountant. During that conversation, I casually remarked that I could probably just build an app for it. To a certain extent, that turned out to be true.

More importantly, it became an opportunity. What started as a simple utility evolved into a focused applied R&D effort exploring how agentic workflows can be used responsibly across the product development lifecycle — from requirements discovery and architectural decision-making through implementation, testing, and deployment.

The resulting Expense Reports App is production-grade and functional. But the real value of this work is what it demonstrates about multi-agent development practices.

The Key Innovation: Using AI Upstream

The breakthrough wasn't using AI to write code — it was using AI upstream of code.

I orchestrated multiple AI agents across the product lifecycle:

  • Draft and refine requirements through conversational brainstorming with ChatGPT
  • Capture architectural trade-offs through ADRs co-authored with AI assistance
  • Translate intent directly into implementation via agentic coding with Codex and Claude
  • Evolve specifications alongside code with documentation agents keeping artifacts current

This context engineering approach — structuring requirements, ADRs, and data models as grounding context for AI agents — reduces the distance between intent and execution. Requirements, decisions, and implementation live together and evolve together. The result: faster iteration with higher confidence.

Grounding in Real Constraints

Early in the process, I grounded requirements in IRS expense reporting and substantiation guidelines. This wasn't just about compliance—it was about disciplined prioritization.

Understanding regulatory requirements clarified: - What elements are universally required - The natural initial market (individuals and small businesses) - Clear MVP boundaries that prevent scope creep

This grounding became a forcing function for product discipline. Rather than building everything that could be useful, I built only what the market demonstrably requires.

AI as a Learning Accelerator

Throughout this work, AI agents served as learning accelerators, not decision substitutes.

When exploring unfamiliar territory — requirements, edge platform capabilities, testing strategies — different agents helped in different ways: - ChatGPT for rapid survey of options and trade-offs on mobile - Claude in Copilot for interactive exploration and supervised refinement - Codex for async implementation from voice prompts, reviewed via PR - Documentation agents for keeping ADRs and release notes current

The key: agentic workflows accelerated my learning and execution, but product judgment remained central. Every architectural decision, every prioritization choice, every trade-off was still mine to make. This is the human-in-the-loop (HITL) pattern — AI agents generate and execute, I validate and direct.

Why This Matters

While the functional domain is expense reporting, the broader value is in the methodology:

  • Multi-agent orchestration across the full product lifecycle, not just implementation
  • R&D conducted through real, production-grade systems with actual constraints
  • MVP discipline informed by reality rather than feature wish lists
  • Context engineering — structured documentation as grounding context for agentic coding
  • Tight feedback loops between product intent and technical execution

This isn't theoretical—it's a working system that demonstrates these practices in action.

Key Takeaways

This applied R&D effort demonstrates several transferable patterns:

  1. Upstream AI Use: Apply agentic workflows to requirements and architecture, not just code
  2. Co-located Artifacts: Keep requirements, decisions, and implementation together as grounding context
  3. Regulatory Grounding: Use external constraints as forcing functions for discipline
  4. Human-in-the-Loop: Leverage AI agents for learning and velocity, not judgment replacement
  5. Hypothesis-Driven: Build real systems to validate approaches, not just prototypes

The result is a production-grade system that serves as an ongoing testbed for agentic development practices. More importantly, these patterns are transferable to any product development context.

Related: For details on what was built, see the Expense Reports App project page.

Try it yourself: https://expensereports.app

Categories: blog

Tags: ai-workflows, devops, system-design