What AI coding agents actually changed (and what they didn’t)
The honest account of building with AI coding agents is not about how fast the code ships. It is about what you still have to do yourself, and how the shape of that work changes when implementation stops being your bottleneck.
Paul Gendek is a Senior Staff Engineer and the creator of Mock Band, a free browser-based backing band app for solo guitarists. He returned to the project in September 2026 after months away. What he found was not that AI did the work. It was that AI reorganized who owned what.
The Original Problem
The idea came from a practical frustration. Jamming alone on guitar meant stopping to reach for a laptop whenever you wanted to change the tempo or groove. With real bandmates, you call for a key change mid-song. Alone, you lose the thread.
In a note from September 17, 2023, Gendek described an app where a guitarist could choose drums, bass, and accompaniment, repeat a groove, and reshape it through voice instructions. A Bluetooth footswitch would trigger voice listening. He named it Mock Band and sketched logo ideas on graph paper.
The earliest preserved repository commit dates to August 2025. The first build happened during parental leave. Repository history shows Cursor use early on, then Copilot later. Gendek describes his role during that period as prompting the implementation rather than writing source code directly. The contributions he claims clearly: concept, platform choice, brand, requirements, and acceptance testing.
️ The Footswitch Showed What Tests Can’t
One of the most instructive moments in the project had nothing to do with code quality. Gendek bought a Bluetooth footswitch to trigger voice input. The implementation assumed hold-to-talk behavior. The physical device worked as a toggle.
No amount of reviewing whether the code looked plausible would have caught that. He had to press the switch, observe what it did, describe the mismatch, and feed that back into Cursor. The interaction got rebuilt around a tap that starts listening while recognition is pending. It required debugging and repeated manual testing before it worked reliably.
That experience made a general principle concrete: a requirement about a physical gesture has to survive contact with the physical object. Software can faithfully implement an assumption that the hardware immediately disproves.
What Changed When Codex Entered the Workflow
When Gendek returned to the project in September 2026, he used Codex to run background tasks in parallel: implementation, automated checks, browser validation, and operational setup. He no longer had to wait for one closely supervised thread to finish before starting the next.
That parallelism sounds like pure upside. It is not quite that simple.
More work arriving in parallel also means more work arriving for review. Some agent-produced fixes introduced regressions. The existence of a change did not settle whether it belonged in the product. Gendek required QA steps in every pull request description and kept human merge approval for himself throughout. When the review queue became unmanageable, he asked for related pull requests and manual checks to be combined into batches.
He is direct about one thing: he did not measure a productivity multiplier, and he will not claim one. The improvement was a review process he could manage, not a quantified speed increase.
The Stack
The current version of Mock Band uses React and TypeScript for its interface and state management. Text input and recognized speech share a command vocabulary. Mapped MIDI actions connect to state and scene handlers. An audio bridge passes changes to an AudioWorklet, which handles drums, bass, rhythm accompaniment, and timing in a separate audio execution context.
Musical changes can queue for beat or bar boundaries. Scenes and preferences persist in the browser, with JSON backup and import available.
On September 29, 2026, Gendek reported a successful live MIDI output test through an IAC bus into Logic Pro. That confirms the IAC route works. It does not confirm that every physical MIDI receiver will work.
What the Review Questions Actually Are
Gendek frames the core discipline this way: the most useful review question is specific. What does this check establish, and what does it leave open?
Some questions can be answered by inspecting software behavior or running a browser check. Some require listening. Physical controller questions require the relevant hardware. Questions about practice value require someone to practice.
Without that boundary, an implementation result becomes a broader product claim by default. That is the failure mode he is trying to avoid.
The Product Decision Agents Couldn’t Make
The original app required an email address before anyone could use it. The list was not growing. Gendek concluded that collecting addresses offered little value and may have discouraged use. In the September refresh, he dropped the gate. Access became immediate. Signup for updates became optional and confirmed.
Agents could implement that change. They could not decide whether it was the right tradeoff. That judgment required knowing what mattered at that stage of the project and what he was willing to give up.
He is careful not to overclaim the outcome: product and distribution changed at the same time, so there is no clean evidence that removing the gate caused growth or improved retention. The early playback data is observational.
The Replication Pattern
If you are using AI agents on a project with physical components or hardware integrations, Gendek’s workflow points to a few specific practices worth copying:
- Require QA steps in every pull request description. Agents produce changes; you decide whether they belong.
- Keep merge approval human. The agent producing a change and the person accepting it into the product are two different roles.
- Batch related PRs when the queue gets out of hand. Parallelism creates review pressure, and batching is a practical release valve.
- Test physical hardware yourself. No automated check substitutes for pressing the actual button.
- Define the scope of each check explicitly. An IAC bus test does not prove MIDI compatibility. State that boundary out loud, or it disappears.
Mock Band is free to use without an account at mock.band/app. Voice and MIDI capabilities depend on browser and device support. Gendek notes that the next milestone for the project is evidence of recurring musical value, not just successful playback.
Disclosure: Paul Gendek created Mock Band and has a vested interest in it. This article is based on his original interview and case study, with AI tools involved in drafting. He directed the product and retained merge approval; agents implemented changes and ran automated checks.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
