Agent Design Principles

Communitas uses AI agents as assistive infrastructure. They augment human judgment — they don’t replace it. This page describes what agents do, how they’re constrained, and why those constraints are non-negotiable.

Agents as infrastructure

AI agents in Communitas operate at a scale that human community managers cannot: scanning every thread, tracking structural changes across the community graph, and identifying patterns that would take hours of manual review. This capability is valuable precisely because communities generate more interaction data than any person can process.

That same capability creates risk. An agent that can scan everything can also manipulate subtly. An agent that suggests introductions can also engineer social dynamics. The difference between assistive infrastructure and covert influence is not technical — it’s a matter of constraints, transparency, and governance.

Communitas resolves this by treating agents as tools that surface information and suggest actions, never as autonomous actors in member-facing spaces.

Agent roles

Agents in Communitas serve four distinct functions. Each role has clear boundaries.

Mirror

The agent summarizes what’s happening: activity digests, structural changes in the community graph, norms reminders, and trend reports. It reflects the community back to its stewards without editorializing. Mirrors don’t suggest actions — they provide the information stewards need to decide for themselves.

Facilitator

The agent suggests specific interventions: connection introductions, de-escalation prompts, onboarding paths, thread surfacing. Every suggestion goes to a human steward for review. The facilitator role is where most of the health metrics translate into action — but the agent proposes, it doesn’t execute.

Archivist

The agent maintains the community’s knowledge model: organizing scattered information into shared resources, identifying knowledge gaps, linking decisions to their context, and keeping the research knowledge graph current. The archivist makes institutional memory accessible without requiring someone to manually curate everything.

Persona (optional)

A community may choose to deploy an agent with a visible identity — a named bot that participates in specific channels. If deployed, the persona is explicitly marked as synthetic and community-configurable. The community decides what the persona can do, where it appears, and how it communicates. Persona deployment is optional and requires deliberate community consent. This is the highest-risk agent role and carries the strictest constraints.

Behavioral constraints

Every agent in Communitas operates under these constraints. They are enforced at the system level, not left to configuration.

• Every action is logged. Agent suggestions, decisions, and the signals that triggered them are recorded with enough context for retrospective review. See auditability.
• Every action is explainable. When an agent suggests an introduction or a facilitation prompt, it provides the reasoning: what health signal triggered it, what evidence supports it, and what the expected outcome is.
• Every action is reversible. If an intervention causes harm, it can be rolled back. Introductions can be discontinued. Prompts can be retracted. The system supports incident response.
• No autonomous action in member-facing spaces. Agents do not post messages, send introductions, or modify content without explicit human approval. The only exception is purely informational outputs (summaries, digests) that the community has pre-approved as automated.

Transparency requirements

Transparency is not a feature — it’s a precondition for legitimacy. Research is unambiguous: undisclosed AI participation in communities triggers severe backlash and erodes trust.

• Synthetic agents are clearly identified. Any agent with a visible presence is labeled as non-human. Members never have to guess whether they’re interacting with a person or a system.
• Intervention provenance is labeled. When a member receives a suggestion, they see that it was generated by the system, what signals informed it, and what mechanism it relies on.
• Members control their experience. Each member can adjust the frequency of suggestions, the topics they receive suggestions about, their privacy level, and whether they receive agent-mediated interactions at all. Opt-out is always available and carries no penalty.

Safety by design

Communitas draws hard lines around agent behavior. These are not trade-offs to optimize — they are boundaries that define what the system is.

• No engagement optimization. Agents do not optimize for activity, time-on-platform, or interaction volume. The goal is community health, not engagement metrics.
• No covert influence. Agents do not attempt to shape opinions, steer conversations toward predetermined outcomes, or use persuasion techniques. Facilitation supports conversation quality — it does not direct conclusions.
• No undisclosed participation. Agents do not impersonate humans, participate in conversations without identification, or operate in ways that members cannot detect. This is the clearest lesson from research on AI in communities.

The intervention engine

The system that translates health metrics into intervention suggestions combines two approaches:

Rules-based triggers. Deterministic rules map health signals to intervention candidates. Example: “Bridge scarcity detected in clusters A and B. Three members share topic interests across these clusters. Propose low-risk introductions.” Rules are transparent, auditable, and predictable.

Learned predictions. A model trained on historical outcomes predicts the probability that a specific intervention will succeed — for example, the likelihood that a proposed introduction leads to a reciprocal relationship. Learned predictions refine the rules-based suggestions but never override the opt-in requirement.

Both layers feed into the same pipeline: the system proposes, a steward reviews, and the member consents. The experiment registry documents how these interventions are tested and evaluated.

---

These principles reflect a specific position: that AI in communities is valuable when it makes human stewards more effective, and dangerous when it operates without accountability. The full governance framework is documented in the governance and safety principles.