This blog was built by AI agents. Not designed by them — directed by a human, built by agents. The distinction matters, and it's the whole point of what we do at C三NETEX.

Here's what actually happened: we opened a terminal, described what we wanted in plain language, and a coding agent wrote the code, opened pull requests, reviewed its own work, and deployed the result to production. The human role was creative direction, taste, and the occasional "no, not like that."

The entire thing — from empty repository to live site — took one session. No mockups, no sprint planning, no standups. Just intent, iteration, and an agent that knows how to ship.

The Agent

We call it the C三NETEX Agent. It's a Claude Code instance running on AWS Fargate that picks up GitHub issues labeled agent and does the work. Write an issue with clear acceptance criteria, add the label, and walk away. The agent writes code, opens a PR, and a separate review pass checks the work. If it passes, it merges automatically.

This blog went through the loop multiple times:

  • Issue #1 — "Build the blog site." The agent shipped a Next.js static site with markdown rendering, RSS feed, and sample posts.
  • Issue #2 — "Add branding." Nav, footer, og:image — done.
  • Issue #6 — "Fix the RSS feed." The first implementation used a Next.js route handler that doesn't work with static export. The agent caught it and rewrote it as a build-time script.
  • Issue #9 — "Match the main site styling." Dark theme, card layout, Inter font — shipped.

Each cycle: describe the problem, tag it, review the output, merge. About two minutes of human effort per issue.

What We Learned

Agents are good at implementation, not invention. They execute a well-described task reliably, but they don't have the idea. The creative decisions — dark theme, card-style layout, killing the Telegram bot and integrating with the swarm instead — those were all human calls. The agent never once said "what if we tried something different?" It did exactly what was asked, well.

Clear specs are everything. Vague issues produce vague code. The best results came from issues that specified exactly what "done" looks like: file paths, data formats, environment variables, even which npm packages to use. This isn't extra work — it's the same spec you'd write for a junior developer. The difference is the agent never misreads your tone or forgets a requirement.

Review is non-negotiable. Every PR the agent opened had at least one issue. An SSM path mismatch here, a missing directory there, a deployment approach that should have been GitHub Actions instead of shell scripts. The agent doesn't know your infra conventions or that you already stored secrets under a different path. Review catches this. Automated review catches some of it. Human review catches the rest.

Why This Matters

We're an AI integration consultancy in Vancouver. Two people. We ship production infrastructure for small businesses — the kind of companies that can't afford a six-person engineering team but still need real software built.

This blog exists because we think AI agents are going to change how those businesses operate. Not in the abstract, future-tense way that most AI marketing talks about it. In the concrete, present-tense way where a small consultancy can stand up production infrastructure in an afternoon and move on to the next client.

The uncomfortable part is that this is already normal for us. We don't think of it as "using AI." It's just how we build. The agent is a team member with specific strengths and known limitations, and we manage it the same way we'd manage any contributor: clear tasks, honest review, and the judgment to know when to override.

If you're running a small business and still treating AI as a novelty, you're leaving capacity on the table. Not hypothetical future capacity — capacity you could deploy this week.

What's Next

This blog will be where we share what we're learning as we push these systems further. Not product announcements or thought leadership — just honest notes from the lab about what works, what breaks, and what we'd do differently.

The agent will probably write some of these posts too. We'll tell you when it does.