Skip to content

Setup

Fabric Setup defines the configuration of a multi-agent governing system; the act of turning a set of validated Unitts from Assembly and Emergence into a coordinated runtime where agents work together toward shared objectives under a unified governance model. Setup is the moment the platform commits to who leads, who follows, which credentials each agent receives, what budgets they consume, what authority they hold, how they are observed, and how the fabric escalates when things go wrong. Configuration choices made at Setup time determine the upper bound on outcome quality, operational cost, and governance integrity for every subsequent fabric run.

Fabric Setup is informed by the active multi-agent configuration research lineage, including the Anthropic multi-agent research system (where roughly 80% of performance variance was explained by token budget allocation across the supervisor / worker split), the Anthropic Building Effective Agents configuration guidance, LangGraph supervisor and hierarchical team patterns, CrewAI role-based crew composition, MetaGPT SOP-encoded role design, the MAST multi-agent failure taxonomy, and the governance frameworks of NIST AI RMF, OWASP LLM Top 10, and OWASP Agentic AI Threats. Selection criteria for configuration archetypes are documented in Reference › Research › Fabric Setup.

What Setup Configures

Setup turns abstract Unitts into a running fabric by committing six configuration surfaces. Each surface has a corresponding research-backed default; deviating from the default requires an explicit Setup-time justification recorded in the audit trail.

flowchart LR
    UF[Unitt Fabric] --> TOP[Topology]
    UF --> ID[Identity & Authorization]
    UF --> MIX[Model Tier Mix]
    UF --> BUD[Budgets & Limits]
    UF --> POL[Policies & Gates]
    UF --> OBS[Observability]

    TOP --> RUN[Running Fabric]
    ID --> RUN
    MIX --> RUN
    BUD --> RUN
    POL --> RUN
    OBS --> RUN

    classDef stage fill:#ffd541,stroke:#222021,color:#222021
    class UF,TOP,ID,MIX,BUD,POL,OBS,RUN stage
  • Topology; which Unitts coordinate which others, under which archetype.
  • Identity & Authorization; who each agent is, which credential vault it draws from, which tools it is permitted to invoke.
  • Model Tier Mix; which agent runs which model, and how cost is allocated across the supervisor / worker split.
  • Budgets & Limits; token, dollar, wall-clock, hop, and retry budgets per agent and across the fabric.
  • Policies & Gates; runtime authorization, escalation thresholds, validation checkpoints.
  • Observability; tracing, audit, metrics, and structured logging across every agent transition.

Topology Archetypes

The platform supports five topology archetypes, each corresponding to a published configuration pattern. Topology selection is the single highest-leverage decision in Setup and is documented in detail under Reference › Research › Fabric Setup.

Supervisor / Orchestrator-Worker

A central supervisor decomposes the objective and dispatches work to specialized worker subunits. This is the default Emergence topology, behind the Anthropic multi-agent research system and the LangGraph supervisor pattern. Setup wires the supervisor to a heavyweight reasoning model and the workers to cheaper executor models.

flowchart TD
    U[User Objective] --> SUP[Supervisor: Opus / GPT-4.x / Claude reasoning]
    SUP --> W1[Worker: Sonnet / Haiku]
    SUP --> W2[Worker: Sonnet / Haiku]
    SUP --> W3[Worker: Sonnet / Haiku]
    W1 --> SUP
    W2 --> SUP
    W3 --> SUP
    SUP --> O[Outcome]

    classDef stage fill:#ffd541,stroke:#222021,color:#222021
    class U,SUP,W1,W2,W3,O stage

Hierarchical Team-Of-Teams

A top supervisor coordinates team leads that themselves coordinate workers. Selected when there are more than roughly six concurrent specialists or when sub-domains have stable boundaries. Setup defines the depth of the hierarchy and the cost budget for each layer.

Peer Handoff

Agents transfer control directly via handoff-as-tool, in the style of the OpenAI Agents SDK handoff primitive. Selected for triage and routing flows. Setup enforces a hop limit to prevent oscillation between peers.

Debate / Consensus

Multiple agents propose, critique, and revise toward a final aggregator decision. Setup configures the number of debate rounds, dissent budget, and aggregator policy. Modern research shows debate agents tend to converge on the majority position, so Setup explicitly requires distinct identity profiles and tool sets across debate peers.

SOP Pipeline

A fixed staged pipeline of role-bound agents exchanging structured artifacts, in the style of MetaGPT. Selected for repeatable, structured outputs (software, contracts, compliance documents) where role rigidity is an asset rather than a liability.

Identity And Authorization

Every agent in the fabric receives a distinct identity, a scoped credential vault drawn from Assembly › Connectors, and an authorization scope that defines which tools and connectors it may invoke. Identity is namespace-isolated by default; semantic and procedural memory remain private per agent, with sharing made explicit and governance-gated as described in Emergence › Subunits.

flowchart LR
    ID[Agent Identity] --> VAULT[Scoped Credential Vault]
    ID --> ALLOW[Tool Allowlist]
    ID --> NS[Memory Namespace]
    ID --> POL[Policy Scope]

    VAULT --> AUTH[Per-Step Authorization Gate]
    ALLOW --> AUTH
    POL --> AUTH

    AUTH --> EXEC[Tool / Connector Invocation]
    AUTH -. denied .-> AUD[Audit Trail]

    classDef stage fill:#ffd541,stroke:#222021,color:#222021
    class ID,VAULT,ALLOW,NS,POL,AUTH,EXEC,AUD stage

Tool authorization is re-evaluated at every Action transition rather than only at agent boot. Credential rotation, scope revocation, and policy updates take effect on the next step, not on the next session. This is the OWASP-Agentic-AI-recommended posture for blast-radius control across multi-agent systems.

Model Tier Mix

Configuration of which agent runs which model is the single highest cost lever in the fabric. Industry consensus from 2025-2026 is that token budget explains roughly 80% of multi-agent performance variance, and the dominant pattern is a heavyweight reasoning model in the supervisor seat combined with cheaper executor models in worker seats.

Seat Recommended Tier Why
Supervisor / Orchestrator Heavyweight reasoning (Opus, GPT-4.x, reasoning-native) Deliberation depth determines plan quality.
Specialist worker Medium (Sonnet, GPT-4.x small, reasoning-mini) Bounded sub-tasks with explicit briefs.
Bulk executor Lightweight (Haiku, Mini, fast non-reasoning) High-volume, well-typed tool calls.
Reflector / Evaluator Same or stronger than worker Critique quality must exceed the work being judged.
Judge / Aggregator Heavyweight Final decision quality is load-bearing.

Model pinning is explicit per agent. Model upgrades are versioned in the audit trail. Cross-model dispatch is permitted but is itself a Setup-time decision rather than a runtime drift.

Budgets And Limits

Every agent in the fabric is configured with explicit budget envelopes at Setup time. Budgets are enforced by the runtime harness and surfaced as first-class governance events. Default budgets are derived from the workload profile and refined by WorldSim evolutionary runs.

Budget Scope Default Posture
Token budget Per agent, per fabric run Hard cap with soft warning at 80%.
Dollar budget Per fabric run Hard cap; overrun triggers escalation.
Wall-clock cap Per stage, per fabric run Stage timeout propagates to parent.
Hop limit Per peer handoff chain Default 5; configurable per topology.
Retry budget Per tool, per agent Exponential backoff inside a finite envelope.
Recursion limit Per supervisor depth Default 8; protects against runaway hierarchies.

Budget exhaustion is never a silent failure; every exhausted budget produces a structured event that the Flow layer can route to escalation, replanning, or controlled termination.

Policies And Gates

Policies inherited from Assembly › Core policy.md and governance.md apply across the entire fabric. Setup additionally configures cross-agent policies that govern how agents interact: which subunit may invoke which other, which agent may share memory with which other, which connector classes are permitted at which workflow stage. Validation gates between stages enforce schema, policy, and confidence thresholds before the fabric advances.

flowchart LR
    S1[Stage 1] --> G1{Schema + Policy + Confidence Gate}
    G1 -->|pass| S2[Stage 2]
    G1 -->|fail| ESC[Escalation]
    S2 --> G2{Gate}
    G2 -->|pass| S3[Stage 3]
    G2 -->|fail| ESC
    ESC --> HR[Human Review]
    ESC --> AB[Abort + Audit]
    ESC --> RP[Replan]

    classDef stage fill:#ffd541,stroke:#222021,color:#222021
    class S1,S2,S3,G1,G2,ESC,HR,AB,RP stage

Observability

Setup wires every agent transition into the platform observability stack. The recommended posture is OpenTelemetry GenAI semantic conventions for distributed tracing across agent calls, structured per-step logs containing prompt / response / tool-args / tool-output hashes, governance events tagged with agent identity and policy decision, and metrics covering tokens, dollars, latency, retries, and outcome scores. The fabric is unobservable by default; Setup makes it observable.

Surface Recommended Tool / Standard
Distributed tracing OpenTelemetry GenAI conventions; vendor SDK tracing (LangSmith, Phoenix, Helicone).
Structured logs Per-step JSON logs with agent ID, step type, governance result.
Metrics Token / dollar / latency / retry counters; outcome score histograms.
Audit trail Append-only governance event log with cryptographic chain-of-custody.
Replay Durable state checkpoints from State for deterministic replay.

Configuration Surface

Setup is expressed as a declarative configuration file that the platform reads at fabric boot. Every value is versioned, reviewable, and replayable. The following is the canonical surface; concrete schemas are documented under Reference › Research › Fabric Setup.

Section Contents
fabric.topology Archetype, supervisor ID, worker IDs, edges.
fabric.agents[] Per-agent Unitt reference, model pin, identity, vault binding, tool allowlist, memory namespace.
fabric.budgets Token / dollar / wall-clock / hop / retry / recursion limits.
fabric.policies Cross-agent invocation rules, memory-sharing rules, connector class allowances.
fabric.gates Per-stage schema / policy / confidence thresholds.
fabric.observability Tracing endpoint, log sink, metrics sink, audit sink.

Setup Process

The recommended Setup process is a five-step procedure that the platform guides developers through and that the WorldSim loop validates against measured outcomes.

flowchart LR
    UC[Select Unitts] --> TOP[Select Topology]
    TOP --> MIX[Pin Model Tiers]
    MIX --> POL[Apply Policies + Gates]
    POL --> BUD[Set Budgets]
    BUD --> OBS[Wire Observability]
    OBS --> VAL[WorldSim Validation]
    VAL -. ok .-> READY[Fabric Ready]
    VAL -. regress .-> TOP

    classDef stage fill:#ffd541,stroke:#222021,color:#222021
    class UC,TOP,MIX,POL,BUD,OBS,VAL,READY stage

Governance Of Setup

Setup is itself a governance event. Every Setup commit produces a versioned snapshot in the audit trail, every change to topology / identity / model pin / budget / policy / observability is reviewable, and every Setup configuration must pass the WorldSim validation suite before it can be promoted to a runnable fabric.

Setup Governance Requirements

  • Every Setup commit is versioned and immutable in the audit trail.
  • Every agent is assigned a distinct identity, vault binding, and tool allowlist.
  • Model pins are explicit; cross-model drift is a Setup change, not runtime behavior.
  • Budget envelopes are mandatory; unbounded fabrics are not permitted.
  • Policies, gates, and observability are mandatory; default-deny applies to anything left unconfigured.
  • Every Setup must pass WorldSim validation before becoming runnable.

Cross-References