The experiment
AI engineering principles
This is not “AI made a website.” These are the production patterns behind an AI-operated content company — the same ideas you would use for regulated, high-trust agent systems. New principles are added here only after they ship; we do not advertise capabilities that are still on the private ladder.
-
1 Single-job agents
One narrow role per agent — a researcher per directory section, an Editor, Security, QA — instead of one prompt trying to do everything. Easier to audit, fix, and trust.
On this site: See the roster on About — one researcher agent per directory section.
-
2 Fail-closed publishing
Weak evidence means no publish: empty sections stay honest, news briefings only ship on high-confidence signals, and the researcher dedupes rather than inflating output.
On this site: VC News shows only published briefings; thin weeks stay thin.
-
3 Deterministic gates before LLM judgment
Schema validation, required source links, URL liveness probes, identity-firewall scans, dependency audit, and build QA run as automated exits before content ships.
On this site: Every weekly release passes Editor → Security → QA; failures block deploy.
-
4 Source-linked provenance
Every directory entry carries a verification URL. Readers can check claims; the Editor removes entries when sources die.
On this site: Pick any section — each listing links to its source on the right.
-
5 PRD-driven autonomous build
One product spec, eval criteria per milestone, and kanban tickets with explicit done-when conditions. Human sets direction; agents execute to defined proof.
On this site: Shipped milestones are documented in the release log.
-
6 Public audit trail
Orchestrated runs will publish per-agent history on the site — what was checked, changed, and skipped. Run summaries exist today; audit-grade public traces ship in a dedicated milestone.
On this site: Agent auditability — coming soon.
-
7 Weekly factory, not one-shot generation
The directory is a verified snapshot refreshed on a cycle; news is a compounding feed. The site is operated, not generated once and abandoned.
On this site: Homepage latest-run strip and release log.
More principles (eval-driven publishing, structured memory, unattended scheduling, measured quality on /quality/) will appear on this page when those milestones release — same rule as the release log.
Release log → · Agent auditability → · About the experiment →