Architecture

Team Design

How MARIA OS designs, composes, and governs multi-agent teams — from role architecture to graduated autonomy.

Foundation

Why Team Design Matters

In autonomous systems, the team is the smallest unit of governance.
How you compose it determines what it can decide — and what it must escalate.

01

Judgment Cannot Be Parallelized

Execution scales linearly. Judgment does not. A team of 10 agents can process 10x the work, but the quality of a single decision depends on how responsibility is structured — not how many actors participate.

02

Composition Determines Capability

The mix of human, agent, and hybrid members defines the team's operational envelope. All-human teams preserve nuance but limit throughput. Agent-heavy teams maximize speed but require stronger governance rails.

03

Roles Before Actors

Design the responsibility structure first, then assign actors. A role defines what decisions can be made, what evidence must be produced, and who bears accountability. The actor — human or agent — is a runtime binding.

04

Trust Is Earned, Not Configured

Teams evolve through maturity stages. An agent earns expanded autonomy by demonstrating consistent, auditable performance within governed boundaries. Configuration sets the ceiling; behavior sets the floor.

Team design is not staffing. It is responsibility architecture.

Composition Spectrum

Watch a Team
Evolve.

From all-human to agent-heavy — 4 tiers of team composition. Each tier activates more of the 8 standard agent roles as governance infrastructure matures.

Planner
Architect
Executor
Quality Eng.
Gatekeeper
Evidence
Doctor
Safety Policy
Chief MARIA
A
All Human
Agents as tools only — human approves every action
Agents: 2/8Throughput: 1x
B
Hybrid Lead
Hybrid coordinator bridges human intent and agent execution
C
Human + Agent
Human sets direction; agents execute within governed bounds
D
Agent Heavy
Autonomous execution — human oversight via fail-closed gates only
Human 100%Agent 0%

Standard oversight

Eight standard roles. Four maturity tiers. Chief MARIA orchestrates them all.

Role Architecture

8 Standard Roles x K1-K8 Skills

Every team draws from the same 8 standard agent roles. Each role maps to the K1-K8 skill model — the operational foundation shared across all universes.

Planner AgentREQ

Plan generation, dependency resolution, execution ordering

K5 PlanK3 Infer
In: Goal definition, Constraints, Context
Out: Execution plan, Resource allocation
Architect AgentREQ

Design decisions, structure definition, interface governance

K5 PlanK4 Verify
In: Plan, Technical constraints
Out: Architecture spec, Interface contracts
Executor AgentREQ

Implementation, optimization, deterministic execution

K6 ExecuteK1 Collect
In: Architecture spec, Requirements
Out: Artifacts, Results
Quality EngineerREQ

Test design, reproducibility verification, quality scoring

K4 VerifyK7 Audit
In: Artifacts, Specifications
Out: Test results, Quality score
Gatekeeper AgentREQ

Boundary verification, permission checks, stop trigger management

K8 GovernK4 Verify
In: Actions, Policies
Out: Approval/Denial, Stop signals
Evidence Agent

Evidence collection, log organization, audit data generation

K1 CollectK2 Curate
In: Execution logs, Decisions
Out: Audit trail, Evidence package
Doctor Agent

Health monitoring, fault detection, self-healing guidance

K3 InferK6 Execute
In: System metrics, Anomaly signals
Out: Health report, Recovery actions
Safety Policy Agent

Safety policy enforcement, risk classification, defense design

K8 GovernK7 Audit
In: Decisions, Risk signals
Out: Safety assessment, Policy updates

K1-K8 Skill Model

K1

Collect

Gather raw data and signals

K2

Curate

Organize and prioritize data

K3

Infer

Derive insights and predictions

K4

Verify

Validate against ground truth

K5

Plan

Sequence actions and allocate resources

K6

Execute

Perform actions deterministically

K7

Audit

Review and score quality

K8

Govern

Enforce constraints and policies

Skills are fetched from the Skill Store or dynamically generated. See /architecture/skills for the full skill lifecycle.

Coordination Topology

Inter-Zone Coordination

Every cross-zone interaction passes through a governance gate. Chief MARIA orchestrates routing via the MARIA Coordinate System.

Hub-Spoke with Governance Gates

4 zones coordinated through Chief MARIA. Dashed lines = gate-governed paths.

Strategy4 agentsExecution5 agentsQuality3 agentsEvidence3 agentsG1G2G3G4
Chief MARIA

5 Gate Types

SchemaReproducibilityGoldenStopObservability

Hub-Spoke

Chief MARIA routes tasks through governance gates to specialized zones. Low inter-zone dependency.

Best for: Independent work streams with shared quality standards.

Coordinate: G1.U1.P*.Z*.A*

Mesh

Direct peer-to-peer coordination between zones. Requires conflict resolution via Gatekeeper agents.

Best for: Tightly coupled domains where zones share state.

Coordinate: G1.U1.P1.Z*↔Z*

Cascade

Linear handoff: Zone A output becomes Zone B input. Each gate validates the handoff evidence bundle.

Best for: Sequential workflows like pipeline processing.

Coordinate: G1.U1.P1.Z1→Z2→Z3

Maturity Model

Graduated Autonomy

Teams evolve through four maturity stages. Each stage unlocks greater autonomy by building the governance infrastructure required to support it.

L1

Manual Oversight

Low risk, low throughput

Team Pattern

All-human teams, agents as tools only

Autonomy Level

None — agents execute explicit instructions

Governance

Human approval on every action

Zero autonomous decisions100% human reviewAgent uptime irrelevant
L2

Assisted Execution

Moderate risk, 3-5x throughput

Team Pattern

Hybrid leads with human oversight

Autonomy Level

Tactical — agents propose, humans approve

Governance

Gate on decisions above threshold

< 30% autonomous decisionsEvidence trails emergingEscalation paths defined
L3

Governed Autonomy

Managed risk, 10-20x throughput

Team Pattern

Agent-led with human quality gates

Autonomy Level

Operational — agents decide within bounds

Governance

Fail-closed gates + exception handling

60-80% autonomous decisionsFull audit coverageValue alignment tracked
L4

Self-Governing Teams

Systematic risk management, 50x+ throughput

Team Pattern

Agent-heavy with governance infrastructure

Autonomy Level

Strategic — agents propose direction changes

Governance

Meta-governance: agents monitor governance health

95%+ autonomous executionSelf-correcting loopsHuman intervention by exception only

More governance enables more autonomy. The ceiling rises with the floor.

Responsibility Distribution

Who Owns the Decision?

Responsibility flows upward through the MARIA Coordinate System: G.U.P.Z.A

GGalaxy (Tenant)

Enterprise boundary — ultimate authority

Owns: Organizational values, ethics file, irreducible constraints

G1 = Bonginkan Corp

G.UUniverse (BU)Galaxy

Business unit scope — strategic direction

Owns: Cross-planet policy, resource allocation, value alignment

G1.U2 = Sales Operations

G.U.PPlanet (Domain)Universe

Functional domain — operational governance

Owns: Domain-specific rules, team composition, workflow design

G1.U2.P3 = Contract Review

G.U.P.ZZone (Ops)Planet

Operational unit — execution boundary

Owns: Task execution quality, SLA compliance, evidence production

G1.U2.P3.Z1 = Auto-Review Zone

G.U.P.Z.AAgentZone

Individual worker — task-level decisions

Owns: Action outcomes, execution traces within assigned parameters

G1.U2.P3.Z1.A5 = Contract Verifier

Core Principles

Responsibility Cannot Be Delegated

An agent executes on behalf of a human, but the human retains accountability. Delegation creates an execution chain, not a responsibility transfer.

Every Decision Has an Owner

No decision exists without a responsible party. If no owner is identifiable, the decision is escalated until one is found.

Evidence Is Mandatory

Every responsibility transition produces an evidence record. The absence of evidence is itself a governance violation.

Escalation Is Not Failure

Escalation is the system working correctly. A team that never escalates is either trivially scoped or dangerously overconfident.

Responsibility is architecture. If you cannot trace it, it does not exist.

Formation Protocol

How Teams Are Formed

9 steps — the same deterministic sequence used by the Universe Builder. Each step produces auditable output.

01

Goal

Define decision scope and business objective

Out: Decision boundary map

02

Scope

Map required roles and escalation paths

03

Team

Bind actors to roles (human/agent/hybrid)

04

Skills

Assign K1-K8 skill categories per role

05

Build

Configure zone graph and coordination topology

06

Gates

Set 5 quality gate types and thresholds

07

Validate

Run constraint checks and conflict detection

08

Test

Execute golden tests and reproducibility checks

09

Deploy

Team is live with telemetry and monitoring

Anti-Patterns

Ghost Accountability

No traceable owner for a decision class. Common when hybrid roles lack clear scope.

Premature Autonomy

Agent-heavy composition without governance infrastructure. Speed up, failures undetected.

Governance Theater

Gates exist but always pass. Evidence produced but never reviewed.

Escalation Avoidance

Teams suppress escalations to appear capable. Looks clean until systemic failure.

Same protocol. Same gates. Same evidence requirements. Every team is a governed universe.