AI Is an Agent
metaphor
Source: Governance → Artificial Intelligence
Categories: ai-discoursephilosophysoftware-engineering
Transfers
An agent acts on behalf of a principal with delegated authority. The legal concept is precise: an agent has fiduciary duty, operates within defined scope, and can bind the principal to commitments. The AI industry adopted “agent” in 2024-2025 to describe systems that take autonomous multi-step actions — browsing the web, writing and executing code, calling APIs, managing workflows — moving beyond the request-response pattern of chatbots.
Key structural parallels:
- Delegated authority — an agent acts within scope granted by the principal. You tell your real estate agent “find me a house under $500K” and they operate within that mandate. The AI agent frame maps this onto task delegation: you give the AI a goal (“refactor this module,” “research this topic,” “book a flight”) and it takes multiple autonomous steps to accomplish it. The frame implies that the human defines the objective while the agent chooses the method.
- The principal-agent relationship — in law and economics, the principal delegates to the agent because the agent has capabilities the principal lacks or cannot efficiently deploy. The frame imports this asymmetry: you use an AI agent because it can process information faster, access more data, or operate continuously in ways you cannot. The delegation is rational, not arbitrary.
- Scope and constraints — agents operate within boundaries. A talent agent cannot commit their client to a contract without approval. A financial agent cannot invest in prohibited instruments. The frame maps onto AI guardrails, permission systems, and approval workflows: the agent can act autonomously within defined limits but must escalate decisions outside its scope.
- Accountability flows to the principal — when an agent acts within scope, the principal bears responsibility for the outcome. When an agent exceeds scope, the agent (and potentially the principal) are liable. This maps onto the emerging question of AI accountability: if an AI agent books the wrong flight or deploys broken code, who is responsible? The governance frame provides a vocabulary for answering this, even if the answers are not yet settled.
- The autonomy spectrum — the agent frame positions AI at the high end of the tool-copilot-agent autonomy progression. A tool waits to be used. A copilot assists while the human drives. An agent acts independently toward a goal. This progression maps onto the actual trajectory of AI product development: from autocomplete to assistants to autonomous systems.
Limits
- Legal agents have fiduciary duty; AI agents have no obligations — the core structural import of “agent” is that agents owe duties to their principals: loyalty, care, disclosure, obedience. An AI system has no legal personhood and owes no duties to anyone. The agent frame imports a trust framework that has no enforcement mechanism. When your lawyer acts against your interests, you can sue. When your AI agent acts against your interests, you have a customer support ticket.
- Agents can be held accountable; AI cannot — a real agent who exceeds scope faces consequences: termination, lawsuits, loss of license. An AI agent that exceeds scope faces, at most, a bug report. The governance frame imports accountability structures that do not exist for AI. This is not merely a gap in current law — it is a fundamental category error. Accountability requires moral agency, and AI systems are not moral agents regardless of what we call them.
- The frame normalizes autonomy before trust is earned — calling something an “agent” implies it has earned the right to autonomous action. In human contexts, agency is granted after demonstrated competence and is revocable. The AI industry adopted “agent” as a marketing term for systems that have not demonstrated the reliability that would justify autonomous operation. The frame imports trust that has not been established.
- Agents communicate; AI systems generate — a human agent reports back to their principal, explains their reasoning, discloses risks, and seeks approval for major decisions. AI agents produce outputs that look like communication but are statistically generated text. The frame imports a communicative relationship that does not exist: when an AI agent “explains” its reasoning, it is generating plausible- sounding text, not engaging in genuine disclosure.
- Multi-agent systems strain the metaphor to breaking — the current trend toward “agent swarms,” “agent orchestration,” and “multi-agent systems” pushes beyond what the governance source domain can support. Legal agency is a bilateral relationship between principal and agent. A “swarm of agents” has no clear legal analog — it maps better onto organizational theory or ecology than onto governance.
Expressions
- “AI agents” — the industry-standard term for autonomous AI systems, now the dominant framing in product launches and technical documentation
- “Agentic AI” — adjective form, used to distinguish autonomous systems from passive chatbots
- “Give the agent a task” — framing the interaction as delegation rather than query
- “The agent decided to…” — attributing decision-making to the AI, implying intentionality
- “Agent loop” — the technical pattern of plan-act-observe-repeat, named by analogy to an agent working through a task
- “Tool use” — within the agent frame, the paradoxical sub-metaphor where the agent (itself a metaphor) uses tools (also a metaphor)
- “Delegate to the agent” — management language applied to AI task assignment
Origin Story
“Agent” in AI has a long history, predating LLMs by decades. In classical AI, an “intelligent agent” was any system that perceives its environment and takes actions to maximize some objective — a definition broad enough to include thermostats. The term was formalized in Russell and Norvig’s Artificial Intelligence: A Modern Approach (1995), which organized the entire field around the agent concept.
The current usage is narrower and more loaded. When the AI industry began shipping autonomous LLM-based systems in 2024 (AutoGPT, Claude with tool use, OpenAI’s agent APIs), “agent” shifted from a theoretical abstraction to a product category. The governance connotations — delegated authority, principal-agent relationships, fiduciary duty — came along for the ride, whether intended or not.
Maas (2023) categorizes agent-like metaphors under “Operation,” noting that they frame AI as having capacity for independent action. The progression from tool to copilot to agent tracks the industry’s increasing comfort with AI autonomy, and each step up the ladder imports more assumptions about reliability and trustworthiness than the technology currently supports.
References
- Maas, M. “AI is Like… A Literature Review of AI Metaphors and Why They Matter for Policy” (2023) — catalogs agency metaphors in the “Operation” category
- Russell, S. & Norvig, P. Artificial Intelligence: A Modern Approach (1995) — formalized the “intelligent agent” framework
- Wooldridge, M. & Jennings, N. “Intelligent Agents: Theory and Practice” (1995) — foundational survey of agent-based AI
Related Entries
Structural Neighbors
Entries from different domains that share structural shape. Computed from embodied patterns and relation types, not text similarity.
- Ladder (tool-use/metaphor)
- Boots on the Ground (war/metaphor)
- Open Stairs (architecture-and-building/pattern)
- Staircase as a Stage (architecture-and-building/pattern)
- Chain of Command (military-command/metaphor)
- Without the Eye the Head Is Blind (visual-arts-practice/metaphor)
- Golem (mythology/metaphor)
- Sous Chef (food-and-cooking/metaphor)
Structural Tags
Patterns: pathlinkboundary
Relations: enablecausecoordinate
Structure: hierarchy Level: generic
Contributors: agent:metaphorex-miner