Agentic Development

I do not think AI is just making developers a little faster.

It changes the shape of software work. The old model was a team moving through tickets, meetings, estimates, and handoffs. My workflow is different: I set the goal, define the boundaries, run agents against the codebase, read the diffs, test the result, and push back until the work is good enough to defend.

The job now

The point is not to protect an old process. The point is to get the right software built, checked, and shipped with clear technical ownership.

Tools I Use

Codex

Repo-wide changes, tests, refactors, documentation updates, and the verification work that turns a draft into a real change.

Claude Code

Deep codebase reasoning, architecture discussion, domain modeling, and long-form work on thornier implementation problems.

Cursor

Fast inline coding, targeted local edits, pairing flow, and focused iteration inside active development work.

What Changed

Agentic development is not just adding a chatbot to the old process. It changes what one technical person can move through in a day.

Normal development

A lot of energy goes into keeping people aligned: tickets, estimates, standups, handoffs, review queues, and release windows. It works, but coordination becomes the job.

Normal AI development

A developer keeps the same process and asks AI for help with snippets, docs, explanations, or local tasks. Useful, but the basic shape of the work has not changed.

Agentic development

I define the outcome, boundaries, and quality bar, then run agents through planning, code, tests, review, and documentation. My job is to direct, challenge, verify, and decide.

Scrum And Agile Are Dead

As default ways to run software, the old rituals no longer match the speed or shape of the work.

Scrum and agile are dead

That sounds dramatic, but I mean it practically: as the default way to run software. Scrum and agile were built around human coordination. AI changed the constraint.

The metaphor changed

Scrum borrowed from rugby. But this is not rugby anymore. A rugby team is not very useful when the other side brought a tank and a swarm of drones.

The new role

The valuable person is no longer just a ticket mover or ceremony keeper. It is someone who can frame the problem, use agents well, verify the work, and make the hard calls.

How The Work Runs

Agents are useful when the work is framed clearly and checked hard. The process is simple, but it is not casual.

Start with the problem

Before I let agents loose in the code, I write down what we are trying to do, what cannot break, and what good looks like.

Give agents clear jobs

Codex, Claude Code, and Cursor each get specific work: implement this, check that, explain the tradeoff, write the test, clean up the edge case.

Make the design argue back

I use AI to challenge the architecture before the team is committed to it. If the design is weak, I want to know early.

Trust the evidence

Tests, diffs, logs, screenshots, and local runs matter more than a confident explanation. If the evidence is weak, the work is not done.

Shape the product faster

Interfaces, flows, copy, data models, and failure states can be explored quickly enough that bad ideas die before they become a roadmap.

Keep the notes honest

Plans, decisions, QA notes, and release context need to move with the code. Otherwise the documentation becomes a fossil.

Principles

AI can write code. It cannot own the consequences.

Agents can do the work, but I still need to be able to defend the architecture, the tradeoffs, and the release.

Planning still matters. Story-point accounting and ceremony matter a lot less when the feedback loop is this fast.

Less process theatre. More working software, better evidence, and clearer technical ownership.

How AI Shows Up In Current Projects

Every current project uses AI in the build process, while product constraints and architecture decisions stay explicit.

FireLog

  • Used agents to move faster through workflow modeling, implementation, debugging, and production polish
  • Kept the compliance details practical instead of letting the tool wander into generic product logic
  • Used AI heavily, but kept the architecture, data, and security calls human-owned

SupportLog

  • Used agents to work through privacy-sensitive states, role boundaries, handoffs, and escalation paths
  • Moved faster on domain modeling, implementation planning, UI iteration, and QA
  • Kept confidential data boundaries and operational visibility decisions under direct review

CrossBorder

  • Used agents to speed up mobile UX iteration, calculator logic, and edge-case checks
  • Kept calculation behavior grounded in explicit domain rules and user-facing estimate language
  • Kept privacy, offline behavior, and degraded modes clear while moving quickly

SpaBooker

  • Used agents to accelerate product shaping, interface iteration, implementation work, and QA passes
  • Kept the workflow grounded in what a small service business would actually need day to day
  • Used AI for speed while retaining direct control over scope, UX, and technical decisions

Maze Generator

  • Used Codex to port the old canvas idea into a typed, static Next.js route while keeping the utility simple and browser-only
  • Used agents for implementation, responsive polish, project wiring, and local verification
  • Kept the scope focused on preserving the original toy as a useful public page

Exorcisor

  • Used agents to move from product idea to working mobile implementation quickly
  • Worked through local-first architecture, data model iteration, UI polish, and HealthKit integration with AI in the loop
  • Kept constraints explicit: offline-first, no accounts, no telemetry, and minimal dependencies