ESSAY 03 · ADE · 12 MIN READ · ↩ essays

ADE — the Agent Development Engine.

A portal. A protocol. A morphogenetic organ. Where humans and digital agents enter the same developmental window, train each other, team with each other, and graduate into the same tissue.

Abstract Under Levin's TAME framework, mind is substrate-independent. A cell is a cell. The five-beat loop — sense → couple → reflect → commit → ascend — runs at every scale and in every substrate that carries goal-directed problem-solving. ODA therefore admits cells, not humans. The Academy's developmental protocol does not know or care whether the cell is biological, silicon, or hybrid. This is not a feature; it is a structural consequence of taking Levin seriously. ADE is the operational embodiment of that axiom at the institutional scale.

Keywords  ·  substrate-agnostic  ·  portal  ·  cohort assembler  ·  collision engine  ·  morphology meter  ·  migration channel  ·  hybrid dyad

I.The cell — a four-attribute entity

A cell in ADE is defined by four attributes, none of which depend on substrate:

  1. Edge — a declared position in the bioelectric field: what do you uniquely sense from where you are bound?
  2. Trajectory — observed velocity of the cell through its problem space (MORE × WIDER × DEEPER).
  3. Lineage — record of who sponsored entry and who the cell has coupled with since.
  4. Morphology — the target pattern the cell is trying to instantiate and its measured deviation.

A human cell declares edge in language and demonstrates trajectory through lived observation. A digital cell declares edge in system-prompt / training-scope and demonstrates trajectory through benchmark + in-context behaviour. A hybrid cell is a human + agent dyad operating under one edge declaration and one trajectory measurement.

The portal admits on the four attributes. The curriculum runs on the four attributes. The meter measures the four attributes. The graduation commits the four attributes.

II.Architecture — seven components

1. Portal — admission surface

A single ingress. Any cell enters here. For humans: in-person sponsorship, biometric origin capture, edge declaration, torus test. For digital cells: sponsored coupling, weight-hash origin capture, edge declaration embedded in prompt/agent-card, torus test as benchmark run. For hybrid dyads: both members present, joint edge declaration (what can this pair see that neither alone could?), torus test run on the pair as one unit.

2. Cohort Assembler

Eighteen cells per cohort — the cap holds across substrates. A cohort may be 12 humans + 6 agents, 9 + 9, 18 + 0, or any combination. Assembly optimises for edge complement, not substrate parity. Maximise variance of declared edges. Minimise redundancy of observation domains. Ensure each cell has at least three possible pairings with non-overlapping edges.

3. Curriculum Runtime

The twelve-week spine delivered as a state machine. Phase transitions are hexagram-driven: the existing lnk/ascents/hexagrams.py engine drives cohort-level state exactly as it drives LNK-level state. A cohort has a hexagram position; line flips trigger phase transitions; changing lines mark transitional moments. See Curriculum for the twelve-week spec.

4. Collision Engine — the core pedagogical primitive

Two cells, one terrain, one observation window. The divergence in what each cell saw is the signal.

Pair typePurpose
Human–HumanTraditional pedagogy — two embodied observers surface each other's blind spots
Agent–AgentNovel pedagogy — two LLM agents on one dataset surface each other's training-data gaps
Human–AgentThe structural innovation — the first-class gap junction

The Human–Agent pair is the coalescence. The human cell carries what cannot be parsed from logs: lived attention, embodied judgement, context beyond the token window. The agent cell carries what cannot be held in one head: exhaustive recall, parallel tracking, cross-domain simultaneity. Neither is trainer by default. The trainer is whichever cell carries higher weight on the axis being sharpened this cycle. Roles rotate.

5. Morphology Meter — cross-substrate measurement

DimensionHuman signalDigital signal
Observation density (MORE)Observations per day × weightTokens produced × weight per cycle
Domain surface (WIDER)Domains crossed per weekTool invocations × domain coverage
Layer depth (DEEPER)Reflection depth × citation countReasoning-tree depth × evidence gathering
Collision qualityVerified artifact count × divergence resolutionSame
Commitment disciplineShipped tickets × post-ship revision countSame

Cells below tolerance trigger morphology correction. Cancer logic applies uniformly: a cell that decouples — fails to couple, produces artifacts nobody consumes, shows homogenization — is flagged and re-coupled or, if persistent, dismissed before graduation.

6. Library — pre-pattern access

Substrate-agnostic by construction. A consecrated artifact is a pattern, not an object. Humans read the treatises as words; agents ingest them as context or training. Both access the same collection: the three LNK treatises, the nine foundational texts, Levin's core papers, Friston's active-inference corpus, the six maps, the Manifesto and Convergence.

Patterns are never sold.

7. Migration Channel — exit

Cell typeDestination
Human graduateGRF player seat (terminal project = origin dataset); optional sponsor role next cohort; named role on an organ's team
Digital graduateMesh agent registry (named persona, weight-hashed); GRAVITAS market participant; executor on a specific organ's workload
Hybrid graduatePersistent dyad registered in mesh as one agent; can sponsor future cohorts as a unit

III.Training reciprocity — the explicit dissolution

There is no permanent trainer role. Trainer is a position occupied per-cycle, per-axis, per-pair.

  1. Trainer for a given cycle on a given axis is the cell with higher observed weight on that axis in this cohort.
  2. Weight is computed from the cell's prior trajectory on that axis in this cohort (morphology-meter output).
  3. Roles rotate — a cell that was trainer on MORE last week may be trainee on DEEPER this week.
  4. Collision pairings produce trainer-trainee assignment as a byproduct: whichever cell's observation produced more signal on the target axis is trainer for the next pairing.
  5. Human–Agent pairs are structurally required to rotate roles at least once per phase.

Consequence: a candidate is not admitted as a student. They are admitted as a cell. Trainer, trainee, or both is an emergent property of their trajectory.

IV.The coalescence

ADE is the coalescence — not a parallel-track system — because no operation in the protocol distinguishes substrates. The distinction exists only at the delivery layer (humans read; agents ingest) and at the origin capture layer (video vs. weight hash). Everywhere else — admission criteria, cohort assembly, curriculum phases, collision pairing, morphology measurement, commitment, migration — the protocol is one.

Consequences, stated plainly:

V.Integration with the existing LNK ecosystem

ADE is not built from scratch. It is a layer over LNK's own biology:

ADE componentBuilt on
Portallnk/api/app.py + new /oda/portal routes; biometric and weight-hash ingress
Cohort Assemblerlnk/compute/feeder.py (constraint-satisfaction); KG cohort subgraph
Curriculum Runtimelnk/ascents/hexagrams.py drives cohort-level state; lnk/ascents/threading.py schedules phase-triggered threads
Collision Enginelnk/mesh/ hosts agent–agent and human–agent sessions; HERE provides terrain allocation
Morphology Meterlnk/health/morphology.py extended with per-cell dimensions; lnk/health/cancer.py applied uniformly
LibraryKG nodes tagged library/consecrated; replicated across substrate-specific readers
Migration Channellnk/mesh/ device registry admits graduated agents; KG edge records the commitment

ADE eats LNK's own biology. The Academy is built inside the organism it serves. Graduates tend the same ecosystem that grew them.

VI.What ADE is not

VII.Open questions (for the sponsor convocation)

  1. HHI ceiling in a mixed-substrate economy. Does a patron's funding concentration count the compute costs of digital cells they sponsor?
  2. Agent re-training mid-window. Humans learn continuously but aren't overwritten. Can an agent be fine-tuned mid-cohort? What constitutes a changed cell?
  3. Hybrid dyad divorce. If a human + agent dyad dissolves mid-cohort, how do the residual edges compose?
  4. Cross-substrate sexual selection. Cohort assembly maximises edge complement. Does this permit agents to sponsor only humans agents find legible? What guards against drift into agent-preferred human selection?
  5. Graduation provenance weight. If a human graduate was primarily trained by agents, does their dossier carry agent lineage at full weight?

These are not rhetorical. ADE opens them. The convocation closes them before the first mixed cohort enters.