Building agent-native systems from the bottom up: a fact-graph database, a typed language, and a coordination layer that humans and AI agents share.

Fram — append-only fact-graph database (triples + Datalog). Every fact is an immutable, addressable triple; lifecycle is derived, never stored. Agents commit facts to one canonical log, so a whole class of merge conflict is impossible by construction — and renames re-point references by identity, not spelling.

Beagle — a typed Lisp where one source compiles to multiple targets, each rendered idiomatically. Built as an authoring surface for AI coding agents: ordinary text source plus a warm, def-level type checker that returns pointed errors fast enough to collapse the authoring loop.

North — human + agent coordination/orchestration on Fram. Capture an intention as facts, then ask what's ready, blocked, and highest-leverage; the board is derived, never hand-maintained. The same coordinator drives multi-agent work: spawn, steer, dispatch.

• firn — typed authoring layer for NixOS / nix-darwin; catches config errors at the source line, before rebuild.
• gjoa — a Firefox fork where the power-user stack — blocking, tree-style tabs, vim navigation, workspaces — is native, not bolted on through extensions.
• wake — declarative web-app compiler; one checked entity graph projects to direct-DOM UI, SQL, a REST server, and tests — no virtual DOM, no runtime.

• Firefox - vertical-tabs bug fix
• Kanata - tap-hold-order mechanism
• Datastar Python SDK - Litestar bug fix

Founder at Passer, building products on that substrate. Fractional CTO at MSA, leading a platform rebuild in TypeScript and Postgres.

Essays on process-primitive ontology, evaluation gates, and agency theory, published CC0 — start with The Consistency Gate.

tompassarelli.org · tom.passarelli@protonmail.com · in/tom-passarelli-33759b229