I'm Devguy. Haagsoft is my one-person studio. I build agentic AI products in type-safe TypeScript and make them reliable: a written goal, evals that measure it and a harness that keeps every run pointed at it.
01The north star
An agent without a north star is just a fast way to be wrong. Agentic reliability starts by writing down whatgood looks like, then building everything else to point at it.
One written sentence of what good looks like, plus constraints and a definition of done, in a spec the agent reads on every run. If I can’t write it down, the agent can’t follow it.
Evals grade each run against the north star: fixed datasets, scored outputs, regressions caught before users see them. A prompt or model change becomes a measured decision.
The harness keeps the agent on course: typed tool contracts, bounded loops, guardrails, retries, tracing and a human checkpoint wherever a wrong move is expensive.
02The loop
Getting an agent to look right is easy. Getting it to be right on the hundredth run is engineering. Every build goes round the same loop, and nothing ships past a step it hasn't passed.
Goal, constraints, definition of done. Short, written, versioned with the code.
Only what the agent needs: the right files, tools and examples, nothing that distracts it.
Small slices the agent can finish and I can read in one sitting.
Tests where state changes, evals where the model decides. Output is a claim until checked.
I read every diff. Autonomy is earned by eval scores, not granted by optimism.
03How I work
Every build starts with what good looks like, written down. The spec, the evals and the harness all point at it, so the agent has something to follow and I have something to measure against.
Generating code and behaviour is cheap now; checking it is the real work. I design for review: small diffs, typed boundaries, tests on state changes and evals on anything the model decides.
Least-privilege tools, schemas the model can’t wriggle out of, bounded loops and explicit stop conditions. Reliability comes from the boundaries, not from a cleverer prompt.
Irreversible or costly actions get an approval step. As an agent’s eval scores hold up over time, the leash gets longer; never before.
Spec-driven slices, semantic HTML first, one language full-stack. Every provider is a rental, so Railway, Fly.io, Vercel, Vultr or Hetzner is chosen because the exit is cheap.
04Active builds
Each build is a testbed for the loop above. Live apps open on their own haagsoft.xyz subdomain; the rest is built in the open.
In progress
An agentic smart-pantry assistant. Snap a shelf or a receipt; vision-powered OCR reads expiry dates and identifies ingredients, then the agent handles expiry alerts, recipe suggestions and an auto-built shopping list.
In progress
Geopolitical signal tracking on top of GDELT. Forecasting and news monitoring that filters the global firehose down to the handful of events that actually move you or your business — including how AI is reshaping international relations.
In progress
A reflection app built around guided questions on a seven-day cycle, with time and date tracking and LLM-assisted answer suggestions when the page stays blank for too long.
Learning exercise
A small GPT built from scratch, following the AlgoMonster LLM course. Not a product — the tax paid so the agentic work above rests on something better than intuition.
Brewing
Quietly in the background. Alongside it sits a deeper backlog — SpecFlow, Agent Observer, Agent Clerk, TotalTidy, Orbit, ThreatMonitor — promoted to active as they are picked up.
05Stack
06Human in the loop
Claude is a collaborator, not autocomplete. It drafts fast; I stay accountable. Every output is a pull request from a quick, confident junior: read it, run it, test it, then merge it.
Reading AINews (Latent Space), The Neuron and TLDR; studying agentic patterns via Matt Pocock and Jack Herrington.
How the apps above were built, what broke, and why the design ended up the way it did.
Developers hand more of the coding to agents and still finish the day exhausted. Here's my current workflow, from an empty GitHub repo and a scratchpad to small agent iterations, and my ongoing journey to find flow.
The new IDEs aren't editors with smarter autocomplete. They're agent workbenches, and the centre of gravity is moving from the editor canvas to the repository, the shell and the evidence an agent leaves behind.
My laptop used to be my development environment. Now I'm building a small system of me, AI agents, tests, evals and persistent machines, and the hard part isn't getting an LLM to write code.