agent

ArchiMate 3.2 Equivalent: Application Component (Autonomous Agent / AI Extension)

Represents an AI agent, autonomous LLM worker, generative copilot, or automated cognitive service operating in the enterprise architecture.

Visual Styling

To apply custom shapes, badges, icons, or CSS to AI agents, use a Style Rule (e.g. style agent where ai_governance.autonomy_level == "Autonomous" { icon "bot" shape "hexagon" }).

---

Syntax

agent <id> "<Display Name>" {
    description "<text>"
    owner "<team>"
    tags ["tag1", "tag2"]
    version "<semver>"
    internet_facing true | false
    model "<model_name>"
    ai_governance {
        autonomy_level "Autonomous" | "Semi-Autonomous" | "Assisted" | "Manual"
        human_in_loop true | false
    }
    cost <amount> <unit>
    risk "High" | "Medium" | "Low"
    criticality "Mission-Critical" | "Business-Critical" | "Operational" | "Administrative" | "Standard"
    time "Tolerate" | "Invest" | "Migrate" | "Eliminate"
    security {
        confidentiality "None" | "Low" | "Medium" | "High" | "Critical"
        integrity "None" | "Low" | "Medium" | "High" | "Critical"
        availability "None" | "Low" | "Medium" | "High" | "Critical"
    }

    relationships { ... }
}

---

Properties

Property	Type	Allowed Values	Description
description	string	Free text	Functional capabilities and prompt/agent goals
owner	string	Free text	AI governance custodian or engineering lead
tags	string[]	Array of strings	Categorization tags (e.g. tags ["genai", "llm-agent"])
version	string	e.g. "1.0.0", "2.5.0"	Software or prompt/model release version
internet_facing	boolean	true | false	Whether agent endpoints/interfaces are accessible from the public internet
model	string	e.g. "gemini-2.5-flash", "gpt-4o"	Underlying foundation model or weights
ai_governance.autonomy_level	string	"Autonomous", "Semi-Autonomous", "Assisted", "Manual"	Scope of autonomous decision-making
cost	cost	cost 500 month or token statement	Inference and API subscription spend
risk	enum	"High", "Medium", "Low"	AI safety, hallucination, or operational risk rating
criticality	enum	"Mission-Critical", "Business-Critical", "Operational", "Administrative", "Standard"	Impact tier
time	enum	"Tolerate", "Invest", "Migrate", "Eliminate"	TIME portfolio classification
ai_governance.human_in_loop	boolean	true | false	Requires human verification before taking external actions
security	block	CIA rating block	Security requirements across Confidentiality, Integrity, and Availability ("None", "Low", "Medium", "High", "Critical")

---

Allowed Relationship Verbs & Targets

Verb	ArchiMate Relation	Allowed Target Entity Types
realizes	Realization	capability, goal, stage, process, value_stream
serves	Serving	actor, process, system, app, agent
reads, writes	Access	data_object
flows_to	Flow	system, app, component, process, data_object, agent
triggers	Triggering	process, agent
influences	Influence	goal, kpi
migrates_to	Transformation	agent
depends_on	Dependency	system, app, component, agent
governed_by	Governance	policy, standard
protects	Protection	actor, system, app, component, infrastructure, data_object, process, agent, trust_boundary

---

Complete Example

agent ai_fraud_sentinel "Real-time Fraud Detection Agent" {
    description "Autonomous cognitive agent inspecting high-frequency payment streams for money laundering patterns"
    version "2.5.0"
    internet_facing true
    model "gemini-2.5-flash"
    ai_governance {
        autonomy_level "Autonomous"
        human_in_loop false
    }
    risk "Medium"
    criticality "Mission-Critical"
    time "Invest"
    security {
        confidentiality "Critical"
        integrity "High"
        availability "Medium"
    }
    tags ["ai-governance", "fraud-detection"]

    relationships {
        this realizes cap_fraud_detection "Automates AML & suspicious activity monitoring"
        this reads data_transaction "Inspects transaction events"
        this protects app_payment_gateway "Monitors API calls for anomalies" { mechanism "Monitoring" }
        this triggers proc_freeze_account "Freezes compromised accounts when confidence > 0.98"
        this governed_by pol_ai_safety "Subject to corporate AI Safety Charter"
    }
}

---

Nesting & Architectural Composition

Agents can be declared globally (at the root level or in a layer/group), or declared inline inside a system or app:

system payment_platform "Core Payment Platform" {
    # Inline Agent (automatically establishes 'payment_platform composed_of fraud_agent')
    agent fraud_agent "Real-time Fraud Agent" {
        time "Invest"
        ai_governance {
            autonomy_level "Autonomous"
        }
    }
}

app payment_gateway "Payment Gateway" {
    # Inline Agent (automatically establishes 'payment_gateway composed_of routing_agent')
    agent routing_agent "Smart Route Optimizer" {
        time "Invest"
    }
}

Composition Rules

1. Single Parent: An agent can only belong to a single parent entity. Declaring or linking multiple composed_of parents to the same agent is a compile error.
2. No Ancestor Relationships: Because a parent container visually encloses its children, an agent cannot have additional direct relationships (e.g. serves, depends_on) to its parent app/system or grandparent system. Model interactions with peer components or external systems instead.