tegwick
48729f6485
Map agent architectures, memory/persistence mechanisms, and approximate-Being catalog onto Soul Frame vocabulary. Write phase-iv synthesis, strengthen H9 on the Agent/APE side, finish SOUL-WP-0005, and mark SOUL-WP-0006 ready.
5.6 KiB
5.6 KiB
Agent Architecture Crosswalk (Phase IV)
Workplan: SOUL-WP-0005-T01
Date: 2026-08-12
Purpose: Map contemporary agent systems onto Body / Mind / Soul / Actor /
Persona / Being / APE without equating product language with ontology.
1. Common industry confusions
| Confused as | Often actually is | Soul Frame correction |
|---|---|---|
| “The model is the agent” | Actor substrate | Actor ≠ Being |
| “Persona prompt = identity” | Persona + maybe Character seed | Persona ≠ Being |
| “Has memory = is a person” | storage feature | memory ⊂ continuity evidence, not personhood |
| “Autonomous = Being” | tool loop / scheduler | autonomy profile ≠ Beinghood (Phase I) |
| “User-facing name = recognition” | UX brand | social recognition needs continuity expectations |
| “Multi-agent team = many Beings” | modular Actors/tools | check constitutive integration level |
2. Architecture families → Soul Frame dimensions
| Family | Typical Body | Typical Mind | Soul / relation | Actor | Persona | Default classification |
|---|---|---|---|---|---|---|
| Stateless LLM chat | chat UI, optional tools | session-only models | none durable | frontier LLM | system/prompt persona | strong APE, not Being (TE-01) |
| Tool-using agent (ReAct / function-calling loops) | APIs, browsers, code runners | planning + tool policies; weak long-term self | usually none | LLM + harness + tools | role prompt (“researcher”) | Functional Agent; APE if person-like |
| Workflow / multi-step orchestrator | pipeline steps, queues | task graph, limited world model | audit logs ≠ Soul | orchestrator + workers | optional role labels | Functional Agent components |
| Generative agents (Park et al. 2023) | sim environment, speech acts | memory stream, reflection, planning | nascent relationship traces in sim | LLM + memory DB + scheduler | character card / seed | APE / proto-Being (RN-001/003) |
| Companion / persistent personal AI | chat, maybe voice/avatar | long-term memory, user model | relationship history with principal | LLM + memory + app backend | branded companion persona | APE → candidate Being if continuity + recognition hold |
| Digital human / avatar | 3D/video body, TTS, lip sync | often thin; dialogue model | presentation surface for relation | rendering stack + LLM | visual persona dominant | Body-heavy APE; Being only if other layers present |
| Enterprise “digital employee” | IT permissions, tickets, email | task policies, RAG | org role, audit trail | agent platform | job-role persona | Functional Agent + social role; rarely constitutive Being |
| Multi-agent swarm | shared tools/channels | distributed planning | collective “identity” marketing | many Actors | team brand | TE-05: profile levels; often not one Being |
| Identity-layer proposals (Qin et al. 2026 etc.) | embodiment hooks | competence models changeable | commitments, signed lifecycle | interchangeable models | governed presentation | Technical approach to Actor ≠ Being |
3. Generative Agents detail (Park et al. 2023)
| Generative Agents component | Soul Frame map |
|---|---|
| Observations | experience into Body/Mind |
| Memory stream | autobiographical store (continuity evidence) |
| Retrieval | Mind access pattern |
| Reflection | model/character formation |
| Planning | agentic Mind |
| Believable behavior | descriptive personhood exposure |
| Social interaction in sandbox | early co-adaptation / relationship traces |
Verdict: excellent APE / proto-Being research system. Does not by itself solve provenance, legitimate identity, constitutive self-maintenance, or Social Beinghood outside the simulation harness (RN-001 Finding 12; RN-003 §13).
Classification: persistent heteronomous individual (architecture-provided creation, scheduling, repair, goals) more than constitutively autonomous Being.
4. Where Actor / Persona / Identity are most often collapsed
4.1 Model-centric products
- Upgrade model → marketed as “new brain” of same companion
- Soul Frame: Actor change; identity depends on continuity profile (H6)
4.2 Prompt-centric personas
- System prompt rewritten → “new personality”
- Soul Frame: Persona/Character change; may be DEGRADED continuity if abrupt
4.3 Account-centric identity
- Login identity treated as Being
- Soul Frame: address continuity anchor only; USURPED risk if process replaced
4.4 Multi-agent “one brand”
- Swarm behind single chat name
- Soul Frame: check constitutive integration (TE-05); may be one Social Persona over many Actors
5. Functional Agent vs Autonomous Agent (Phase I reminder)
| Functional Agent | Toward Autonomous Individual | |
|---|---|---|
| Question | What should it do? | Does it self-individuate and self-maintain? |
| Typical stack | tools + goals from outside | + persistence, self-repair, viability conditions |
| Personhood | optional | still separate axis |
| Beinghood | not implied | still needs personhood + address + recognition |
6. Bibliography anchors
| ID / cite | Role |
|---|---|
| Park et al. 2023 (SF-R016) | generative agents architecture |
| Qin et al. 2026 (SF-R043) | governable individual / identity layer |
| Otsuka et al. 2026 (SF-R044) | AI identity standards/gaps |
| Tallam 2026 layered mutability | persistent self-modifying agents |
| Classic agent literature (Wooldridge etc., pointer) | BDI / autonomous agents tradition |
| ReAct / tool-agent surveys (pointer) | contemporary LLM agent patterns |
Full entries: sources/bibliography.md § I; foundations 001/004 references.