How we work
The methodology below isn't a brochure — each practice was forged on a delivered system and is cited in the corresponding case study.
A lean human team plus an AI-agent fleet
Agentryx delivers founder-led, with an integrated fleet of AI agents covering full-stack engineering, QA automation, DevOps and project management. On a live government platform this model replaced a conventional 5–8 person team — that delivery is our proof point, documented in the HireStream case study. The human owns vision, trade-offs and final approval; agents produce artifacts plus evidence, never unverified claims.
Two-gate completion
Nothing is “done” because an AI (or a human) says it is done. A change is complete when the automated checks pass and the behaviour is verified in a real browser — both gates, every time. This rule is enforced on our client platforms and on this website itself.
Five-layer embedded testing
Our standard testing architecture, prototyped in production delivery: (1) unit and integration suites against a real database, (2) deep-smoke route and authorisation checks, (3) synthetic monitoring heartbeats, (4) log mining with LLM-assisted triage, and (5) a per-change confidence score that gates merges.
Evidence beats invention
Every project we deliver is documented in a structured profile: what was built, the architecture and its rationale, the hard problems, and outcomes labelled measured or projected. Those profiles are the only source this website is allowed to draw from — the build fails if a page cites something the evidence corpus doesn't contain.
Release engineering as a first-class product
Air-gapped installers with rollback recipes, four-environment pipelines with mandatory staging dry-runs, production safety guardrails that hard-block dangerous commands. On systems where an outage is a public-service outage, the release process is part of the product.
Configuration over code
Statutory fees, seasonal windows, policy toggles — values that change by government order live in the database, not in code, so policy changes never need a deployment. The same discipline gives operators self-service controls instead of terminal access.
This website practises what it preaches.
Every change to agentryx.dev ships through typecheck, lint, unit tests, browser smoke tests on desktop and mobile viewports, and Lighthouse performance/accessibility budgets — the same two-gate discipline we apply to client work. A content-provenance gate rejects any page that makes a claim the project evidence doesn't support.
See the evidence →