Building software with AI agents

Role: Sole engineer, directing AI agents
Year: 2026
Stack: TypeScript, NestJS, Flutter
Status: Ongoing

I build large systems with a team of AI agents and run it like an engineering team: each agent has one role, every change is reviewed, and tests run before anything is merged.

WorkspaceSprintTasksConcept

Tasks

RoleTaskAssigneeStatusProgressUpdated

SupervisorSplit the sprint into tasksSupervisorPlanning1m
BuilderPayments endpointBuilder 1Building2m
BuilderReminders serviceBuilder 2Waiting for review4m
BuilderBooking screenBuilder 3Building3m
ReviewerReview: booking screenReviewerApproved6m
BuilderProfile settingsBuilder 1Approved9m
ReviewerReview: reminders serviceReviewerWaiting for review4m
BuilderSearch filtersBuilder 2Approved12m
SupervisorPlan the next sprintSupervisorPlanning14m

Why I work this way

AI agents write code fast, but fast code is not the same as reliable code. So I use the structure of a normal engineering team: one session plans and reviews, specialized agents build, and nothing is merged before review and tests.

The team

A lead session runs several agents in parallel, each with one role: agents that build, an agent that reviews, and agents for simple tasks. Every piece of work gets an independent review before merge.

Rules written into the project

The rules live in the project, not in my head. Hooks block unsafe actions automatically, project skills describe how each type of task is done, and a decision log records the reason behind each decision. The log now holds more than 300 decisions.

Tests

More than 6,000 unit tests and 2,400 integration tests, with full coverage on code that handles money. Every change runs against them before merge.