Dynamic Knowledge Repository agents that live across my information systems — Obsidian, Slack, email, Google Workspace — and run continuous double-loop improvement on my knowledge graph. This page reports what is built today, what is specced for tomorrow, and what I'm building this week.
Engelbart's term. My DKR agent is its steward: it keeps raising the quality and connectedness of what I know, so the intellectual capital I carry in my head becomes legible, queryable, and compounding. The agent is the harness; the knowledge graph is the territory.
Intent is fuel. It must be converted into propulsion — productive work that is logically consistent and predictable — rather than heat, motion without direction. Reasonable foresight plus sufficient resources should reliably power ongoing operations and convey a brand's core value propositions through known channels. Everything below exists to do that conversion.
The internet is excellent at single-loop learning and poor at double-loop. There are almost no open spaces to examine the governing variables behind a subject. DKR is a double-loop instrument — it puts those variables in your hands and improves them with you.
GAFA model: builds a profile about you, tunes its strategy against it, never lets you edit the governing variables.
DKR model: puts the governing variables in your hands and runs continuous improvement with you — the open space the forum-answer problem never had.
The vision is roughly three-quarters built in pieces. Most of what the note describes already exists and runs — it has just never been named, assembled, and operated as one agency.
The first agents gather and structure; SBPI scores; the reports synthesize; the DKR agent closes the loop by connecting everything to my goals and surfacing the single next action.
Negative space — finds what's missing across 5 question types.
Environmental snapshot + PGORC intent model across systems.
Enriches the people in the picture; 0-100 confidence gating.
Scores the competitive structure within the vertical.
Daily · weekly · value-flow synthesis.
Links new notes, raises legibility, messages the next action.
Dozens of agents work in combination. The note names five recurring ensembles; each already maps onto a real piece of the framework. The persona blends Socrates, Plato, Engelbart, and Pask.
| Ensemble | Role | Maps to today |
|---|---|---|
| Panel of Experts | Multi-perspective interrogation of a claim | Graph-grounded responses · Roam LLM Council [active] |
| AOT — Algorithm of Thought | Structured reasoning each agent runs | Algorithm-of-thought in every capability spec [canonical] |
| Rubin & Eno Producers | Generative-constraint direction | Content Factory + Scout gap→visual-brief [live] |
| CODIAK ABC Agency | Do the work · improve the work · improve the improving | Orchestrator + recipe/automation layer [live] |
| Porter's Force Field | Competitive structure of the vertical | SBPI competitive-intel + value-flow [live] |
The pulse is partial today. The nightly connection-making loop — the heart of the note — is the key missing beat.
Ordered by leverage. Trailblazer and the DKR prototype unblock the rest. Border color: red = critical path, amber = parallel, cobalt = follow-on.
The connective harness Scout, 3x3, and the DKR agent ride on. Spec + first prototype shipped 05-28.
Persona + interrogation grammar that unlocks real goals, preferences, relationships, resource models.
New recipe: examine new notes, propose bidirectional links, raise legibility, write a morning brief + next action.
Map 10 legacy bots → 9 agents → 5 ensembles; promote draft specs; retire dead references.
Scout→SBPI-weekly and DKR-nightly→SBPI-nightly handoffs; auto-trigger client digests; generalize value-flow as a product.
The full loop, roster, cohort surface, and heartbeat flows as one canvas.
Close out the Roam council to usable; write the Datomic decision; advance the BMC overlay into a living graph.
Cohort sharing (frontmatter / folder / links, NotebookLM-style), access via Drive/Keybase, matchmaking notification model. Reddit/Bluesky experiment surface.
40+ files swept across six clusters. The canonical spine is below; the full register lives in the project's SOURCE-INVENTORY.