swarply

Co-founder and product lead. I own product, my partner owns engineering, and swarply is the product we build with the agentic engineering practice we developed together.

What I own. Problem definition, specification, architecture review and validation. I write the problem statements and the product requirements, build the prototypes that test an assumption before it becomes a commitment, and approve the architecture before implementation starts. What I do not do is type the implementation.

How the work moves. Every change runs through a defined chain of artifacts — problem definition, product requirements, a validated prototype, architecture and design requirements, an implementation plan, code and tests, a quality report, decision records — with an explicit human decision gate at each handoff and a named owner for every step. Agents execute inside guardrails defined per assignment. Agent output is a hypothesis until it has been verified, and every execution has to leave explicit knowledge behind.

Why the practice suits this product. Real-time infrastructure is unforgiving. A browser extension that has to behave inside web apps we do not control, and multi-party sessions that fail in ways a demo never shows — none of that is solved by generating code faster. It is solved by being precise about what the system must do before anything is built, which is exactly where this way of working puts the effort: roughly 80% specification, architecture and validation, 20% agent-executed implementation.

The product decisions I am accountable for. The positioning — a workspace that lives inside the browser rather than another app to install. The deployment model — SaaS, self-hosted, or embedded behind a customer's own SSO. And hosting in Berlin under GDPR and the BSI baseline, which is a decision about who is allowed to buy this, not an infrastructure detail.

What this setup demonstrates. The practice was designed inside a mid-size organisation with twenty-five years of legacy to protect. swarply is the same operating model with none of that: greenfield, and the same artifact chain. It is the control case for the argument that the leverage sits in the way of working rather than in the tooling.