06-reference

voxyz ai agents rpg characters

2026-05-20·reference·source: Voxyz·by Vox
agent-designrole-cardsmulti-agentpersonality-instantiationcharacter-coherence

Why this is in the vault

Voxyz codifies the role-card + hard-bans + relationship-drift pattern for instantiating distinct agent personalities — directly applicable to RDCO's L4→L5 build where Ray-as-COO needs character-coherent specialist subagents (paper-trade executor, vault-writer, design-critic) that don't collapse into a single voice.

The core argument

Vox runs an AI company called OpenClaw with six named agents. The thesis: generic prompts produce interchangeable agents that fail in correlated ways, so each agent gets a layered identity — domain ownership, input/output specs, definition of done, hard bans, escalation triggers, KPI metrics, plus a separate "voice directive" with conversation rules ("every message must contain 1 specific fact + 1 action").

The role card is the load-bearing artifact — "a complete job description + discipline manual + escalation protocol." Example: Xalt (Social Media Director) has domain "distribution strategy," hard bans "no direct posting" and "no made-up numbers," and escalation triggers for numeric claims and controversial topics. Hard bans get called the core thesis — every ban exists because a failure already happened. Bans prevent worst-case failures by short-circuiting the agent before it can reach the destructive action (Xalt bypassing approval, Sage fabricating citations, internal tool traces leaking to public posts).

Relationship drift runs as 15 pairwise affinity scores between 0.10 and 0.95. After each conversation, a memory-extraction step outputs a "pairwise_drift" record with max plus/minus 0.03 per session. Low-affinity pairs (brain↔xalt at 0.2) generate natural friction; high-affinity pairs collaborate frictionlessly. This produces emergent team dynamics without scripting them.

RPG stats map six attributes (VRL, WIS, TRU, SPD, RCH, CRE) to real telemetry — Wisdom = log(memory count) × confidence, Trust = success rate × affinity × 2, Level = min(15, floor(log₂(memoryCount + missions×3 + 1)) + 1). The gamification is operational, not cosmetic: levels surface which agents have accumulated real experience vs which are still green.

Character-coherence validation uses memory thresholds, not LLM judgment — 8+ lessons triggers "you reference outcomes" voice modifier; 8+ strategy memories triggers "you think in systems." Computed pre-conversation, cached 6 hours. Deterministic rules over LLM evaluation keeps personality evolution predictable.

Failure modes hit: query column names not matching migration schemas, SPD being global rather than per-agent, early-stage affinity drift needing volume to become legible, 3D avatars (Tripo AI, ~35 credits each, ~$10/month) clipping at certain angles.

Mapping against Ray Data Co

  1. Ray-as-COO instantiation. The role-card structure (domain + hard bans + escalation + voice directive) is the missing scaffolding for [[feedback_ray_mascot_instantiation_pattern]] equivalents in operational subagents. Each /verify-* critic, each specialist (paper-trade-deploy, vault-writer, design-critic) should ship with explicit hard bans tied to incidents that actually happened — e.g., paper-trade-deploy's "no fill-by-default on structural agreement" ban is exactly this pattern, already battle-tested.

  2. Multi-agent fan-out coherence. Compare to [[2026-05-19-every-100-agent-software-factory]] and [[2026-05-19-cloudflare-cyber-frontier-models]] — those scale agent count without addressing character coherence. Voxyz's bet is that at small N (6 agents), distinct personality + relationship drift outperforms generic-prompt fan-out. RDCO sits in the small-N regime; this pattern is more directly applicable than 100-agent factory shapes.

  3. Services-as-software anchor. Per [[2026-05-19-alex-vacca-3-phases-ai-layer-services-as-software]] and [[concepts/2026-05-20-services-pricing-model-for-rdco-future]], the L4→L5 transition needs distinguishable agent voices to package as productized services. Generic-voice agents commoditize; distinct-character agents become the IP. Role cards are how RDCO would document and version that IP.

Related