Putting agentic software development into practice: Lessons from a real-world project
Flor Cabello · 3 November 2026
Read the originalModus Create rebuilt a shop's mobile app in five months with coding agents, moving to nearly one designer per developer and requiring an engineer's approval of each plan before code.
Flor Cabello, a product leader at Modus Create, a software consultancy, described rebuilding the mobile app of a large online shop whose users rarely returned after their first session. The client wanted the new app before Black Friday, five months away, for work that would normally take more than a year. The team used Claude Code with written instructions, called Skills, that set the points where an engineer must decide.
- The team moved close to one designer per developer, against the usual one for every three or four, because developers working with AI coding agents moved faster.
- A planning Skill made the agent propose what it would change and how, and an engineer approved, edited or rejected that plan before any code was written.
- Before opening a pull request, the engineer compared a screenshot of each screen from an iOS emulator with the approved Figma design, so mismatches never reached testers.
- The engineer who submitted a pull request had to understand and defend the code, a tech lead reviewed it, and testers checked the feature against the original Jira ticket.
A team handing implementation to coding agents can fix in advance the points where a person approves the plan, checks the screen against the design and owns the merged code. Cabello reports that testing and code review then carried more load as pull requests grew. This is a consultancy's account of its own delivery method, and it gives few measured results.
Before opening a pull request, have the coding agent render each new screen in the mobile emulator, and compare the screenshot with the approved design.
When a feature's behaviour is hard to show in static screens, build an interactive prototype in code with the real design tokens and let stakeholders try it on their phones.
Derived by Working Surface from the article; more in the Playbook. Source line: What it does not acquire at any stage is the authority to approve its own work.
6 November 2026: teams scaling AI agents are replacing case-by-case checking with one repeatable path that every piece of agent work passes through.