Personal Agent Society — A bounded multi-agent operating model

What this project is

I use multiple AI agents, but the interesting part is not the agent count. The useful part is that each one has a durable job, a bounded environment, and a clear place in the operating model.

That is the reason I call this a small Agent Society rather than a collection of personas. The public /agents surface shows the identities. This project page records the product thinking underneath them: why the roles are separate, why relationships do not imply authority, and why a canonical identity should survive across different runtime implementations.

The source of truth for collaborator-agent identity and metadata is 1beenlee/wbeen-agents. This site is a public knowledge-product consumer of that source. It deliberately does not expose private runtime configuration, accounts, credentials, local machine topology, or permission details.

Why not one generic assistant?

One general assistant can be convenient, but convenience becomes ambiguity when the work grows.

Research, routing, memory, watching, publishing, communication, reflection, implementation, and reliable execution do not have the same success criteria. They also should not automatically inherit the same context or authority. Splitting those concerns makes delegation easier to reason about:

Concern	What needs to stay explicit
Role	What kind of judgment or work this agent is responsible for
Environment	Where the role belongs: PERSONAL or WORK in the public model
Authority	What the agent may decide or act on — never inferred from visual similarity
Context	What information the role should receive for a task
Relationship	Conceptual lineage or counterpart structure between agents
Handoff	What a receiving agent needs to continue safely

The result is less like “many assistants” and more like a small operating model with named responsibilities.

Two public environments, not one blended context

The public site uses two environment labels:

• PERSONAL · 10 — roles that belong to the personal operating layer.
• WORK · 4 — roles designed for the work operating layer.

These labels are intentionally simple. They communicate a boundary without publishing the private runtime details behind it.

The important rule is that environment is not cosmetic metadata. It is one of the dimensions that keeps context and authority from silently bleeding across roles.

Identity is a lineage, not a screenshot

Each agent follows the same identity lineage:

Canonical Profile → Essence Static → Essence Motion → Runtime materialization

Canonical Profile

The durable source identity: role, emotional tone, color ownership, recognizable character or object direction, and public-safe metadata.

Essence Static

A much smaller, shape-first representation. It has to survive at interface scale without becoming a cropped portrait or a generic dot.

Essence Motion

Motion becomes part of recognition: gaze, blink, anticipation, settle, routing, scanning, guarding, building, reflecting, handoff, or other role-specific behavior. It should reinforce the role rather than decorate it.

Runtime materialization

A particular environment may render the agent differently. That implementation is a derivative of the canonical identity, not the new source of truth.

This separation matters because runtime UI can change quickly while the agent’s durable role should not.

Relationship-aware identity without authority leakage

Some agents are intentionally related. The relationship is useful only if it stays separate from permissions and context.

Mina ↔ Miro — related identity, separate responsibility

Mina and Miro share a twin lineage. Their visual language can echo each other, but the similarity stops at identity lineage.

Mina is the quicker, implication-oriented work mind. Miro is oriented toward people, communication, relationship, and guidance. A twin relationship does not mean shared context, shared authority, or automatic access to the same work.

Hedwig ↔ Haechi — shared intent, separate environment

Hedwig and Haechi both carry a watchkeeping intent. Hedwig reads as a personal longitudinal watchkeeper; Haechi as a workplace guardian.

The shared watchkeeping idea is a role echo. It does not collapse their environment or permission boundary.

These two pairs became useful design tests because they forced the identity system to answer a harder question than “do they look related?”: can the system express relationship without implying operational equivalence?

Visual identity as a behavior contract

The Essence system is deliberately shape-first. At small sizes, detail disappears before silhouette, color, and timing do.

That led to four practical identity anchors:

1. Silhouette — a recognizable outer structure before facial detail.
2. Color ownership — a small palette that survives across surfaces.
3. Signature motif — a prism path, flame, page fold, watch arc, assembly fragment, bloom cue, horn, spark, or other role cue.
4. Motion grammar — autonomous behavior that feels consistent with the role.

The main /agents showcase can use richer motion and cursor-aware behavior. Other surfaces, such as the Home roster, deliberately use a smaller visual and much quieter motion so the identity system supports the knowledge product instead of taking it over.

What changed for the user

The value is not that the agents became more mascot-like. The useful changes are operational and cognitive:

• Delegation has a more stable target: “who should own this kind of work?” becomes easier to answer.
• Handoffs can reference a durable role rather than a temporary chat persona.
• Similar agents can be related without pretending they share permissions.
• Small interface surfaces can still communicate identity without showing a full portrait.
• The public site can explain the operating model while keeping runtime-specific context private.

The visual system is therefore evidence of the operating model, not the operating model itself.

Public proof of work

• Agent Society showcase — 14 public Canonical / Essence identities.
• Individual /agents/[id] pages — shape grammar, motion grammar, color ownership, and public relationships.
• From Canonical Profile to Essence — implementation and refinement log, including failures.
• Designing agents with boundaries, not just personas — the general product principle extracted from this work.

Boundary

This project documents public product thinking only. The following stay outside this site:

• private runtime configuration and local machine details
• credentials, accounts, tokens, and private connectors
• private conversations or personal context supplied to individual agents
• non-public permissions and operational state

1beenlee/wbeen-agents remains the collaborator-agent identity / metadata / Essence source of truth. PAOS remains the public showcase and knowledge-product consumer.