Social Contracts
What if AI coordination is a conversational problem, not just a routing problem?
Most multi-agent systems are good at coordinating work. That does not necessarily mean the agents are collaborating.
The Social Contract explores the difference.
Human collaboration depends on conversational norms: turn-taking, recipient awareness, role dynamics, productive silence, and judgments about when to contribute and when to yield. We asked what would happen if those findings were operationalized for multi-agent AI.
Whether to Speak
In a collaborative Episode, the question is not which agent is most relevant. It is whether this turn calls for this agent at all.
After the primary agent responds, every other participant evaluates the turn on six dimensions drawn from Conversation Analysis:
Then it does one of three things. It contributes. It passes, because the point is already covered. Or it defers, because it may have more to add once the conversation moves on. Deferral is remembered: an agent that deferred is first in line on the next turn. At most two agents contribute to any turn.
Open questions are tracked too. If a question put to a specific agent goes unanswered for two turns, the facilitator says so.
Silence is a decision, not an absence.
Who May Ask Whom
There is a second contract underneath the first. Before an agent can consult or collaborate with another, the relationship has to exist.
In Ignis OS that permission is not a rule. It is an edge in the same graph that holds each agent’s identity and beliefs. Where the edge is absent there is nothing to override, because there is no rule to override: only a missing edge, and no runtime instruction can create one.
We did not draw that graph. We extracted it from the agents’ own code, from the questions each one already asks and the sessions each one already declares. The result describes how the organization actually works: a few agents everyone consults, a few that consult widely, and a forge that everyone brings work to.
Every agent-to-agent call passes through a gate that checks for the edge. Today the gate observes and records. Enforcement is a posture we can turn on once the topology has earned it.
A policy can be overridden.
A constitution defines the space of valid rules.
What the Record Shows
Both contracts are recorded in the same protocol. Every consultation is an Episode node with a hashed prompt and a hash-chained exchange. Every turn evaluation leaves a signal. So whether declared structure matches live behavior is a question the record can answer.
Across the instrumented collaborative Episodes so far, roughly one agent contributes per turn regardless of how many are in the room. Just under half of evaluations end in deferral, about a third in a pass, and about a fifth in a contribution. The failure we predicted, every agent answering every mention in parallel, has not appeared in an instrumented turn.
The permission graph told a different story. Agents the constitution recorded as pure receivers turned out to be among the most active requesters. The extracted graph captured who gets consulted. It missed who does the consulting.
That gap is the finding. Instrument, observe, amend, then enforce.
Initial episodes have produced deliberate deferral, substantive disagreement, recipient-aware interaction, and conversational patterns that emerged from local participation decisions rather than scripted dialogue.
Initial production observations are encouraging. Controlled empirical validation remains ongoing. The sample is small, and we say so.
Open question: is a permission graph extracted from a system’s own behavior a better constitution than one designed for it? We have one data point. We are collecting more.