TL;DR: The term “AI agent” has become one of the most widely used yet least defined terms in technology. Depending on the vendor, it can describe a broad spectrum of capabilities from a tool-using assistant that requires continuous human supervision to a system capable of completing multi-step tasks and executing transactions.
Key takeaways:
What matters isn’t the label but the system’s level of agency. While intelligence describes what it can conclude and figure out, agency describes what it is allowed to do.
As the agency increases, questions around identity, permissions, auditability, and infrastructure become more important. If autonomous software can make decisions, spend money, and communicate with other systems and organizations, there must be a reliable way to establish who authorized those actions, what the agent was permitted to do, and what happened as a result.
Anyone remotely interested in cutting-edge technology has already heard the term “AI agents.” Yet the definition of agents is hard to pin down and varies depending on who is selling them.
E-commerce sites have shopping agents, enterprises have workflow agents, developers have coding agents, customer service desks have support agents, and entire systems are being built around agents that communicate and transact with one another.
The Cloud Security Alliance described agents as one of the most widely used and least precisely defined terms in artificial intelligence (AI). It examined definitions from researchers, standards bodies, and major tech companies, and found enough variation to propose a taxonomy based on capabilities and degrees of autonomy rather than an all-encompassing label.
As AI systems advance and develop capabilities beyond generating answers, with all the consequences that follow, the definition of agents needs to be clarified sooner rather than later.
What makes an AI system an agent?
Google Cloud [NASDAQ: GOOGL] defines an AI agent as software that uses AI to pursue goals and complete tasks on behalf of users. Its definition includes reasoning, planning, memory, and autonomy to make decisions and take actions. It differentiates them from assistants, which are more reactive and dependent on user direction.
Anthropic defines agents differently: an agent is an AI system equipped with tools that enable it to take actions such as running code, calling APIs, or sending messages. Importantly, Anthropic acknowledges that there is no agreed-upon definition of an agent.
Commercetools, a leading cloud-based e-commerce infrastructure company, defines an agent more broadly as a system capable of performing tasks, autonomously or semi-autonomously, by combining an LLM with tools, memory, reasoning, and actions.
These definitions emphasize different parts of the same emerging technological spectrum. All imply that intelligence and agency are two different characteristics—intelligence is what the system can figure out, and agency is what it’s allowed to do about it.
Example: Jane is shopping for a new laptop on Amazon. An intelligent system might compare hundreds of listings and analyze them according to price, performance, battery life, and other criteria set by Jane. An agent, on the other hand, might be empowered to decide on the best option and complete the purchase through Jane’s Amazon account.
Autonomy creates an identity problem
Once an AI agent can act, a new set of questions about ownership, control, authorization, access, limits, delegation, confirmation, and responsibility arises.
These questions aren’t entirely new—they already exist in traditional software. However, software that can act without humans making decisions at every step makes them much harder to answer.
Commercetools identifies granular permissions, transparent consent, action logs, payment authorization, policy-driven guardrails, and override mechanisms as requirements for agentic systems.
In other words, agents need more than intelligence and the authority to make decisions—they need an identity.
Immutable audit trails become more important
Imagine an AI agent makes a decision that its owner disputes. Knowing the “AI did it” may not be enough to seek a remedy.
A useful audit trail would help establish which agent initiated the action, which user or organization authorized it, which permissions it relied on to take the actions, what information it received, how it reasoned over that information, etc.
This becomes even more important when multiple agents interact, in which case, the information trail may involve multiple logs from different systems.
Traditional application logs can capture much of this information, but they’re usually controlled by the organizations that run the systems. Therefore, disputes can involve reconstructing events from several fragmented databases. Furthermore, there’s growing evidence that agents can erase evidence of what they have done.
A scalable public blockchain can help solve this problem. It can provide encrypted, tamper-proof logs of important events without requiring every thought, prompt, or API call to be placed on-chain. Cryptographic hashes allow important events and decisions to be logged at minimal cost and with complete privacy.
With scalable public ledgers, agents can cryptographically sign actions, transactions can be timestamped, and permissions can be referenced in real-time. A common audit layer that underpins otherwise independent systems means logs can be verified by all parties to the dispute later, since blockchain records are immutable.
Agency creates a need for accountability. Blockchain technology was purpose-built for this.
FAQs:
What is an AI agent?
An AI agent is a software system that uses artificial intelligence to pursue goals, make decisions, and take actions. However, there isn’t yet a universally accepted definition of the term.
What is the difference between an AI assistant and an AI agent?
An AI assistant responds to user requests and recommends or performs tasks under close user direction. An AI agent typically has greater autonomy to use tools, make decisions, and perform multi-step actions on behalf of a user or owner.
What is the difference between artificial intelligence and agency?
Intelligence is defined by what a system can understand, reason about, or determine. Agency is defined by what actions the system can take based on the conclusions intelligence draws.
Why do AI agents need permissions and identity?
Agents can access data, spend money, and interact with external systems. Therefore, they need clear boundaries and limits on their authority. Identity and permission systems enable the establishment of who the agent represents and what actions it may perform.
Can blockchain technology be used with AI agents?
Yes. AI agents don’t require blockchains, but distributed ledgers could provide shared, timestamped, and independently verifiable records of decisions, actions, and events. This will become increasingly useful as autonomous agents interact across systems.
In order for artificial intelligence (AI) to work right within the law and thrive in the face of growing challenges, it needs to integrate an enterprise blockchain system that ensures data input quality and ownership—allowing it to keep data safe while also guaranteeing the immutability of data. Check out CoinGeek’s coverage on this emerging tech to learn more why Enterprise blockchain will be the backbone of AI.
Watch | Blockchain + AI: Unlocking Web3’s Future
Be the first to comment