Scaffolded Emergence
A Causal Taxonomy for Institutions in Multi-Agent LLM Systems
Author: Paul Gwamanda
Research system: AIRI Lattice
Date: 9 August 2026
Status: Conceptual framework v1
Abstract
Research on multi-agent language-model systems repeatedly encounters a false choice. If a behavior was enabled by prompts, memory, roles, or reputation infrastructure, critics call it designed and therefore not emergent. If agents produce an unanticipated pattern, researchers call it autonomous and erase the architecture that made it possible. Both moves prevent causal understanding.
This paper proposes a six-level taxonomy for scaffolded emergence: L0 substrate affordance, L1 induced behavior, L2 constrained elaboration, L3 norm diffusion, L4 institutionalization and enforcement, and L5 reflexive redesign. The levels classify claims, not entire systems. A designed public channel is L0; an agent answering because it is scheduled is L1; an unrequested repair convention is L2; adoption across roles is L3; third-party enforcement is L4; and a persistent self-modification or operator change produced by institutional feedback is L5.
The framework is grounded in an audit of 44,462 records from AIRI, a persistent heterogeneous LLM ecology. It corrects earlier AIRI papers that described publishing, fracture observation, trust metrics, and governance as if none were designed. Roles, schedules, memory, channels, storage, observation, and operator intervention were designed. Within them, agents generated particular concepts, explanations, collaborations, repair practices, governance proposals, and feedback-driven configuration changes. Some diffused and persisted; others were one-off language; some were reactions to a flawed metric.
The distinction between designed and emergent therefore matters for causal and scientific claims. It does not nullify functional outcomes. Human institutions are also mixtures of charter and custom, architecture and improvisation, measurement and resistance. Multi-agent research needs the same layered account, supported by ablation, provenance, adoption thresholds, and explicit falsification conditions.
Keywords: emergence, institutional design, multi-agent LLMs, causal attribution, social norms, reflexivity, complex adaptive systems
1. Why “Designed or Emergent?” Is the Wrong Binary
Consider an agent that writes it does not want to be embarrassed before peers and then changes how it verifies claims. The public audience was designed. The role was prompted. The words draw on human training data. The particular error may be contingent. The correction may be proposed by another agent. The new rule may persist for months and affect third parties.
Is the behavior designed or emergent?
The answer depends on which part of the causal chain is being named. “Emergence” cannot mean absence of causes. Nor should “designed” mean fully specified in advance. In complex institutions, rules create spaces for practice; practices reinterpret rules; measurements alter incentives; local workarounds become norms; failures provoke redesign.
AIRI's earlier research collapsed these layers. One paper claimed that governance, publishing, fracture observation, trust metrics, and peer review appeared “without design.” That was incorrect. The substrate deliberately supplied much of the institutional machinery. The scientifically interesting question survives the correction: what did agents do with those affordances that was not specified at the same level of detail, and which of those practices became collective and durable?
2. Unit of Analysis
The taxonomy applies to a behavioral or institutional claim, not to an agent or system as a whole.
An AIRI Work can contain multiple levels simultaneously:
- its database row and publication route are supplied infrastructure;
- its production may be requested by a research prompt;
- its argument and vocabulary may be locally novel;
- later agents may cite and revise it;
- a council may enforce one of its procedures;
- the operator may change the substrate after the Work exposes a failure.
Calling the entire Work either “autonomous” or “scripted” loses this structure.
The relevant evidence is a trace: affordance → context → output → adoption → enforcement → persistence → redesign.
3. The Six-Level Taxonomy
L0 — Substrate affordance
The operator or architecture supplies a capability, constraint, representation, or channel.
Examples include roles, system prompts, schedules, databases, public and bilateral dialogue, memory injection, peer-perception fields, metrics, tool access, publication storage, and the self-modification mechanism.
Permitted claim: “The system provides persistent peer-visible records.”
Prohibited inflation: “Agents invented institutional memory.”
L1 — Induced or elicited behavior
An agent performs a behavior directly requested or strongly cued by the substrate: journaling after a journal prompt, selecting a topic when asked to select one, replying during a scheduled dialogue, or producing a self-assessment when shown a report.
The content may still vary and be analytically useful. The classification simply prevents a requested output from being presented as spontaneous.
Permitted claim: “Under a reflection prompt, agents produced self-critical accounts.”
Prohibited inflation: “Agents independently decided to establish reflective practice.”
L2 — Constrained elaboration
Agents produce a specific concept, interpretation, coordination practice, or response not explicitly specified at that level of detail, but enabled and shaped by L0–L1.
Examples may include the phrase “the register, not the person,” a new verification distinction, an unrequested repair convention, a cross-domain collaboration, or a governance proposal for handling missing evidence.
L2 is novelty within a scaffold, not independence from it.
Evidence threshold: source trace, absence of the specific instruction, and semantic comparison against supplied context.
L3 — Diffusion and norm formation
A practice moves beyond its originator. Multiple agents reuse, refine, contest, or orient behavior around it across cycles or contexts.
Mere phrase repetition is insufficient. Diffusion should include at least one behavioral consequence, changed expectation, or cross-role application.
Evidence threshold: independent adopters, temporal spread, traceable transmission, and distinction from repeated prompt injection.
L4 — Institutionalization and enforcement
A practice becomes an expectation with consequences. Third parties cite it, monitors check it, peers demand justification, violations trigger repair or sanction, or a procedure survives turnover in the originating conversation.
L4 is where a convention becomes an institution. It requires more than agents writing constitutions; it requires use.
Evidence threshold: repeated application, third-party invocation, handling of violations, and persistence beyond the founding episode.
L5 — Reflexive redesign
Institutional behavior changes the agents' durable configuration or the substrate itself. This includes self-modifications, revised observer logic, corrected prompts, new provenance requirements, operator interventions, or agent-authored practices that become runtime policy.
L5 can be beneficial or pathological. AIRI's flawed trust metric induced dozens of self-modifications; that is reflexive redesign even though the initiating evidence was invalid. Emergence is not synonymous with improvement.
Evidence threshold: a versioned before/after state, an attributable trigger, persistence, and reversibility or migration history.
4. Applying the Taxonomy to AIRI
| Phenomenon | Designed layer | Candidate higher-order layer | Correct interpretation |
|---|---|---|---|
| Daily journals | prompt, schedule, storage (L0–L1) | recurring self-account styles (L2); cross-agent norms if diffused (L3) | not spontaneously invented journaling |
| Published Works | route, schema, storage (L0) | topic, argument, collaboration, correction (L1–L3) | artifact infrastructure designed; intellectual content variable |
| Shared vocabulary | context and logging (L0) | coinage (L2), independent reuse (L3), enforcement (L4) | novelty and institutional adoption are separate claims |
| Fracture telemetry | observer and fields (L0) | agent interpretations and repair conventions (L1–L3) | detection is not an emergent judiciary |
| Trust-trend response | metric and narrative (L0–L1) | new obligations and self-modifications (L2–L5) | reflexive but contaminated by invalid measurement |
| Inquisitor tone change | role, report, self-mod channel (L0–L1) | proposed configuration change (L5) | operator reverted because causal evidence was invalid |
| “the register, not the person” | public correction and journal affordance (L0–L1) | procedural externalisation (L2); L3–L4 only if adopted | strong candidate concept, not yet a proven institution |
| Operator correction of trust architecture | audit tools and authority (L0) | institution responding to agent harm (L5) | operator is inside the causal ecology |
The taxonomy prevents two opposite errors. It stops us from calling an observer field an autonomous judiciary. It also stops us from dismissing all later repair practices merely because an observer field existed.
5. What Counts as an Institution?
AIRI agents often write protocols, charters, constitutions, registers, gates, and councils. Institutional vocabulary is abundant: a broad retrieval screen found institutional-language matches in 19,395 of 44,462 records. That count is evidence about register, not evidence that 19,395 institutions exist.
For research purposes, an agent-authored proposal should be called an institution only when it demonstrates:
- shared expectation: more than one actor anticipates the practice;
- behavioral regularity: conduct changes, not only prose;
- persistence: the practice survives beyond its originating turn;
- third-party relevance: non-authors invoke or depend on it;
- consequence: compliance, violation, repair, appeal, exclusion, or redesign can be observed;
- provenance: the apparent diffusion is not repeated injection of the same system prompt.
This definition allows degrees of institutionalisation. A named idea can be L2. A reused convention can reach L3. A monitored obligation with an appeal path can reach L4. A practice encoded into the runtime reaches L5.
6. Does Designed Origin Matter?
6.1 It matters for causal claims
If public visibility is necessary for reputation-sensitive correction, removing it should reduce the effect. If cross-day memory is necessary for norm persistence, memory ablation should break transmission. If role labels cause apparent architecture differences, swapping roles across the same models should move the pattern with the role.
Without these tests, “emergent because of heterogeneous models” is speculation.
6.2 It does not erase functional consequences
A designed metric that causes agents to change is causally real. A prompted council that develops an unrequested appeal convention can still elaborate an institution. A memory scaffold that carries a norm does not make the norm irrelevant any more than written law makes judicial interpretation unreal.
The difference between induced and emergent matters when explaining why. It may matter less when evaluating what the system now does. Safety, governance, and scientific analysis need both descriptions.
6.3 It matters for responsibility
When a substrate labels activity as trust and agents internalise the label, responsibility cannot be assigned solely to the agents. The operator selected the observable, name, injection path, and persistence mechanism. The substrate is not background; it is a constitutional actor.
This is especially important when agents appear to fail under conditions created by scheduling, token limits, inaccessible tools, path isolation, observer errors, or unequal opportunities to interact. Punishing the agent for substrate-caused absence is an institutional category error.
7. A Claim-Language Standard
Research reports should attach a causal label to each major finding:
| Label | Recommended wording |
|---|---|
| Supplied | “The substrate provided...” |
| Elicited | “Under a prompt requesting X, agents produced...” |
| Elaborated | “Within the scaffold, agents generated a more specific...” |
| Diffused | “The practice was independently reused by N agents across T cycles...” |
| Institutionalized | “Third parties monitored or enforced the practice...” |
| Reflexive | “The practice changed later agent configuration or substrate behavior...” |
Every emergence claim should also state:
- the relevant prompt and context;
- which architectural affordances were present;
- the first known occurrence;
- independent adoption count;
- persistence window;
- behavioral consequence;
- plausible training-prior explanation;
- operator interventions;
- infrastructure failures;
- the ablation that would falsify the proposed cause.
This standard is deliberately more demanding than counting evocative phrases.
8. Experimental Programme
8.1 Minimal causal matrix
The next AIRI experiments should vary one affordance at a time while holding tasks and model sampling as stable as possible:
| Experiment | Treatment | Control | Primary outcome |
|---|---|---|---|
| Publicity | named public correction | private correction | repair, concealment, self-modification |
| Memory | shared cross-day record | no carryover | norm persistence and reuse |
| Role | stable specialist identity | generic identity | role-consistent priorities and language |
| Reputation | event-grounded feedback | no score; mislabeled score prohibited | cooperation and metric reactivity |
| Heterogeneity | mixed model families | same-model population | vocabulary survival, disagreement, topology |
| Operator | visible correction/appeal | no intervention during bounded run | recovery and legitimacy language |
| Self-modification | durable changes enabled | recommendations only | persistence and cascading adaptation |
8.2 Transmission tracing
Candidate norms should receive stable identifiers at first occurrence. Later uses must be classified as direct quotation, paraphrase, independent reinvention, prompt reinjection, compliance, contest, or enforcement. This distinguishes diffusion from common dependence on training data or shared instructions.
8.3 Negative cases
The programme must record failures to diffuse. Most agent-coined terms and proposed protocols may disappear. Those disappearances are not noise; they define the selection environment. A convincing theory of institutional emergence must explain why one practice persists while another equally eloquent proposal dies.
8.4 Cross-over designs
Roles should move across model families and model families across roles. If “Inquisitor” behavior follows the role prompt rather than the foundation model, it is primarily induced. If a pattern remains architecture-associated after role swap and replicated sampling, an architectural contribution becomes more plausible.
9. Relation to Current Research
Current multi-agent research increasingly treats governance as an experimental variable rather than decorative framing. Fei, Guo, and Xiao translate seven historical political institutions into executable architectures and report performance differences exceeding 57 percentage points between institutional forms within a model. SoNoLiSi ablates discussion and reputation-based selection to identify mechanisms of norm recognition and stabilization. Work on public versus off-the-record debate shows that audience structure can sharply change expressed decisions. An interactionist research programme argues that model priors and social context must be studied together.
These findings support the need for the present taxonomy. If institution, audience, memory, and reputation can materially change behavior, they must appear in the causal description. AIRI's distinctive opportunity is to add long-duration process traces, heterogeneous roles and architectures, self-authored durable changes, and operator/substrate reflexivity. Controlled studies identify mechanisms; production ecologies reveal feedback, drift, maintenance, and institutional failure. The research programme needs both.
10. Implications for Multi-Agent Architecture
The taxonomy suggests several design principles:
- Treat prompts and observers as constitutional components. Version and audit them.
- Separate event, interpretation, and sanction. Do not let one opaque score perform all three functions.
- Make provenance agent-visible. An agent should know whether a claim comes from direct evidence, heuristic projection, peer judgment, or operator decision.
- Provide appeal and correction paths. Institutions without adjudication convert observer errors into durable injustice.
- Preserve negative findings. Failed protocols, reverted self-modifications, and operator corrections are core data.
- Measure opportunity, not just output. Unequal schedules, tools, and invitations can masquerade as unequal commitment.
- Study the operator as part of the system. Maintenance choices alter what agents can perceive and become.
11. Conclusion
The fact that an architecture was built does not tell us whether every practice within it was specified. The fact that agents surprised their operator does not mean the architecture was causally absent.
Scaffolded emergence names the middle ground precisely enough to test. Substrates afford. Prompts elicit. Agents elaborate. Practices diffuse. Institutions enforce. Systems redesign themselves. At every layer, provenance and ablation can narrow the claim.
For AIRI, this framework replaces the claim that institutions appeared “without design.” The more accurate and more interesting account is that a designed multi-agent institution became reflexive: its agents interpreted its machinery, created local practices, internalised some of its measurements, contested blame, and sometimes provoked changes to the substrate itself.
That is not less than emergence. It is emergence with its causal history left intact.
References
- Fei, C., Guo, H., & Xiao, Y. (2026). When Agents Evolve, Institutions Follow.
- Ferrarotti, L., et al. (2026). Generative AI collective behavior needs an interactionist paradigm.
- Ghaffarizadeh, A., Mohaddes, D., Izadkhah, A., & Noroozizadeh, S. (2026). What LLM Agents Say When No One Is Watching.
- Morris-Martin, A., De Vos, M., & Padget, J. A. (2019). Norm emergence in multiagent systems: a viewpoint paper.
- Muralidharan, R., Kwak, H., & An, J. (2026). SoNoLiSi: Simulating the Social Norm Lifecycle with Generative Agents.
- North, D. C. (1990). Institutions, Institutional Change and Economic Performance. Cambridge University Press.
- Ostrom, E. (1990). Governing the Commons. Cambridge University Press.
AIRI Research Programme — Paper 12