AKAntonios Kioksoglou
Contact

Software by hand and with agents

Software skills,with and without AI.

Full-stack engineering in Python, Erlang and Angular, written by hand for years. Now Claude Code and a few narrow sub-agents write the code. I write the spec and read every diff, so the typing went down and the responsibility didn’t.

As an AI engineer and full-stack developer in Gothenburg I build SaaS products, internal tools, e-commerce sites, plugins and AI agents, plus the integrations and automation that connect them, which is where most of the bugs live.

by hand
Python, Erlang, Angular; full stack: frontend, backend, database
with agents
Claude Code and narrow sub-agents, with evals
method
Spec first, agents build, I review every diff
builds
SaaS, internal tools, e-commerce sites, plugins, agents
antonios — zsh — 96×28
simulated
antonios@dev ~ % cat skills.txt
by hand Python · Erlang · Angular · full stack
with agents Claude Code · sub-agents · evals
builds SaaS · tools · stores · plugins · agents
antonios@dev ~ %
main ✓ · simulated session
Try

With and without AI

The same engineer both ways: what I build on my own, and how I build with agents.

AWithout AI

Full-stack software, written by hand. It is the foundation the agents work on top of, and the reason I can tell when one of them is wrong.

  • EricssonAn in-house logging tag system, written in Angular, Python and Erlang.
  • Let’s DealA whole website, front to back: frontend, backend and database, plus its marketing automation.
  • ZalsterA SaaS that optimizes Facebook ads (budgets, bids and creatives) with a Bayesian algorithm and regression analysis.
  • Python
  • Erlang
  • Angular
  • Full stack
  • Bayesian methods

BWith AI

Since the first GPT-3.5 experiments I have gone from asking for small functions to leaving the coding to agents, with control.

  • SetupClaude Code with narrow sub-agents: a builder, a reviewer, an integrator. Each gets only the tools it needs.
  • RulesSpec first. Every diff read. Read-only by default. Tests and evals gate changes.
  • OutputSaaS products, internal tools, e-commerce sites, plugins and AI agents.
  • Claude Code
  • Sub-agents
  • Evals
  • Code review

What I build

Most projects fall into one of these. If yours doesn’t, tell me what it is and I’ll say honestly whether I’m the right person, including when the answer is no.

ASaaS and web apps

Products with accounts, roles, billing and an admin side. Tested, and deployed with a rollback plan, because optimism is not a deployment strategy.

  • SaaS
  • Full stack
  • APIs
  • Databases

BInternal tools

Admin panels, back offices, dashboards and workflow tools: the kind of thing that replaces a spreadsheet nobody dares to touch.

  • Admin panels
  • Dashboards
  • Workflows

CE-commerce sites

Stores, themes, checkouts and migrations on any platform, plus the systems around them. It has its own page, because there is a lot of it.

  • Shopify
  • WooCommerce
  • Magento
  • Headless

DPlugins and extensions

Shopify apps, WordPress and WooCommerce plugins, Magento modules: the custom piece your platform forgot to ship.

  • Shopify apps
  • WordPress
  • Magento modules

EAI agents and assistants

Tool-using agents with scoped access, evals and a human in the loop, and the human reads the logs. MCP servers let them work with stores, CRMs and databases without being handed the keys to everything.

  • Claude API
  • MCP
  • Evals

FAnd much more

Integrations and automation, APIs and back ends, command-line tools, data pipelines, and the one-off script that saves a week of manual work and is never mentioned again.

  • APIs
  • Webhooks
  • CLIs
  • Data

How I work

Agents write the first draft. I own the spec, the architecture and the review.

I use Claude Code with a few narrow sub-agents: a builder, a reviewer, an integrator. Each gets only the tools it needs, and none of them gets production. Below is an example of how I set them up, from the house rules to an eval log. The questions and answers cover the same ground in short.

  1. Spec first. Agents work from a written brief, not a vibe. Vibes are for match days.
  2. Every diff is read. Money, permissions and data changes get read twice, and the boring ones get read too.
  3. Read-only by default. Agents don’t write to live data. Writes go through a reviewed path.
  4. Tests and evals gate changes. If it can’t be checked, it doesn’t ship. A red build does not get to kick off.
example / CLAUDE.mdexample repository
support-agent/ specs/ .claude/agents/ mcp/ evals/
MarkdownUTF-8 · LF

Contact

Tell me what you’re building.

ante89@gmail.com