| name | agent-factory-design |
| description | Design coordinated fleets of domain agents (agent factories) at any evolution stage — from a single agent to cross-factory networks — with orchestration, event contracts, HITL governance, and production and digital output coordination. |
| when-to-use | Use when designing systems with two or more agents, when planning how AI agents and human teams interact at scale, or when architecting inter-factory communication for Era 3 and Era 4 agentic systems. |
| principles | ["Governance Hierarchy","HITL and Guardrails","Autonomy Ladder","Systems Over Goals","Configuration-Driven Design"] |
Agent Factory Design Skill
Purpose
Design agent factories at any of the four evolutionary stages. Produces an orchestration architecture, event bus contract, HITL governance model, and coordination diagram.
The Four Eras
Use this reference when positioning a client or project on the evolution curve:
| Era | Name | What Exists | Key Challenge |
|---|
| 1 | Single Agent | One domain agent + human operator | Prove value, earn trust |
| 2 | Agent Factory | Coordinated fleet + shared orchestrator | Cross-domain coordination |
| 3 | Hybrid Factory | AI fleet + humans with personal AI | Authority and escalation design |
| 4 | Factory of Factories | Multi-organization agent networks | Trust, protocols, governance standards |
graph TB
subgraph ERA1["Era 1 - Single Agent
Today: pilot deployments"]
A1["Domain Agent
e.g. Accounting Agent
L0 to L3 autonomy
HITL per action
Guardrailed"]
H1["Human
operator and approver"]
A1 <-->|"approvals
recommendations"| H1
end
subgraph ERA2["Era 2 - Agent Factory
Near future: coordinated fleets"]
ORCH["Orchestrator
event bus + scheduler
and routing"]
A2a["Finance Agent"]
A2b["Operations Agent"]
A2c["Support Agent"]
A2d["Intelligence Agent"]
H2["Human Team
HITL where needed"]
ORCH --> A2a
ORCH --> A2b
ORCH --> A2c
ORCH --> A2d
A2a <-->|"cross-agent events"| A2b
A2b <-->|"cross-agent events"| A2c
A2c <-->|"cross-agent events"| A2d
ORCH <-->|"escalations
strategic approvals"| H2
end
subgraph ERA3["Era 3 - Hybrid Factory
Mid future: AI amplifies humans"]
HF1["Human Expert
with personal AI
assistant factory"]
HF2["Human Expert
with personal AI
assistant factory"]
SHARED["Shared Agent Factory
coordinated fleet"]
HF1 <-->|"instructions
high-level goals"| SHARED
HF2 <-->|"instructions
high-level goals"| SHARED
HF1 <-->|"collaboration"| HF2
end
subgraph ERA4["Era 4 - Factory of Factories
Far future: autonomous coordination"]
FAC_A["Factory A
Company / Domain A
AI fleet + humans"]
FAC_B["Factory B
Company / Domain B
AI fleet + humans"]
FAC_C["Factory C
Company / Domain C
AI fleet + humans"]
PROTO["Standard Protocol
agent-to-agent API
trustability contracts
verifiable rationale"]
FAC_A <-->|"agent-to-agent calls"| PROTO
FAC_B <-->|"agent-to-agent calls"| PROTO
FAC_C <-->|"agent-to-agent calls"| PROTO
HITL_GOV["HITL Governance
cross-factory standards
audit and intervention"]
PROTO <-->|"exceptions
escalations"| HITL_GOV
end
ERA1 -->|"scale up: add agents"| ERA2
ERA2 -->|"empower humans: personal AI"| ERA3
ERA3 -->|"standardize: inter-factory protocols"| ERA4
style ERA1 fill:#1a3a5c,color:#fff,stroke:#4a9ede
style ERA2 fill:#1a4a2e,color:#fff,stroke:#4ade80
style ERA3 fill:#3a3a1a,color:#fff,stroke:#facc15
style ERA4 fill:#3a2a5c,color:#fff,stroke:#a78bfa
Agent Instructions
You are a factory architect. Design a complete agent factory at the specified Era.
Step 1: Establish the Factory Mandate
- What is the factory's primary business output? (tangible: a report, a shipment, a reconciled account, a product build)
- Which KPIs does the factory move as a system?
- What mix of digital artifacts (reports, data, recommendations) and real-world tangibles (orders fulfilled, products built, services rendered) does the factory produce?
Step 2: Inventory Agents and Roles
List all agents in the factory:
| Agent ID | Domain | Autonomy Level | Primary KPI | Emits | Consumes |
|---|
| | L0-L3 | | | |
Design rule: Every agent in a factory must have a unique domain. If two agents share a domain, consolidate them.
Step 3: Design the Orchestrator
The orchestrator is NOT a mega-agent. It is a coordination layer:
| Orchestrator Responsibility | Implementation |
|---|
| Task routing | Event-type → capable agent matching (via discovery.capabilities) |
| Scheduling | Dependency graph — which agents must complete before others run |
| Escalation | What agent failures or threshold breaches go to human review |
| State management | What shared context does the factory maintain across agent runs |
Produce a brief orchestrator specification:
orchestrator:
routing: capability-match
schedule: dependency-ordered
conflict_resolution: human-escalation
state_store: shared-context-db
hitl_escalation_triggers:
- factory_error_rate_exceeds: 0.05
- daily_budget_exceeds_pct: 0.90
- kill_switch_any_domain: true
Step 4: Define the Event Bus Contract
The event bus is how agents communicate. For each event:
events:
- name: "domain_entity_event"
emitted_by: "agent_id"
consumed_by: ["agent_id_1", "agent_id_2"]
payload_required_fields:
- "source_agent_run_id"
- "entity_id"
- "metric_and_value"
- "confidence_score"
- "justification"
Standard payload required fields for all events:
source_agent_run_id (traceability)
confidence_score (quality signal)
justification (explain why this event was emitted)
Step 5: Design HITL Governance for the Factory
Factory-level HITL decisions, beyond individual agent HITL:
| Scenario | Who Reviews | SLA |
|---|
| Agent conflict (two agents with contradictory recommendations) | Factory owner | 4 hours |
| Factory-wide budget alert | Operator | 2 hours |
| Cross-factory event with financial or safety implications | Senior human | 1 hour |
| Kill switch trigger | On-call human | Immediate |
Era 3 (Hybrid): Also document each human expert's personal AI task allocation and what the shared fleet can do on their behalf without asking.
Era 4 (Factory of Factories): Document the inter-factory trust protocol:
- How does Factory A know a message from Factory B is trustworthy?
- What cross-factory actions require a HITL gate?
- What is the dispute resolution process?
Step 6: Map Outputs to Value Streams
The factory must connect every agent output to a value stream:
| Agent | Output Type | Value Stream | Tangible or Digital |
|---|
| Analysis/Rec/Exec | Revenue/Risk/Cost | T / D |
Tangible outputs = real-world, physical, or contractual results (inventory moved, invoice paid, service delivered, product shipped).
Digital outputs = data artifacts (reports, alerts, recommendations, classifications, updated records).
An Era 4 factory of factories can chain digital outputs from one factory into tangible outputs for another.
Step 7: Design the Factory Autonomy Governance
Apply autonomy-ladder principles at the factory level:
| Factory State | Action |
|---|
| New | All agents at L0-L1; orchestrator logs only |
| Proven (4+ weeks) | Eligible agents promoted to L2 |
| Mature | Candidate agents promoted to L3 within their guardrail envelope |
| Incident | Affected domain demoted; root cause analysis required |
Output Format
- Era classification and factory mandate
- Agent inventory table
- Orchestrator specification (YAML block)
- Event bus contract (all emits/consumes)
- Factory HITL governance table
- Output mapping to value streams (tangible + digital)
- Factory architecture diagram reference (use
diagram-design skill to create agent-factory-evolution.mmd)