Technology Overview

For decades, artificial intelligence in retail meant systems that assisted human decision-making. Recommendation engines suggested products, but merchants chose which to display. Fraud detection flagged suspicious transactions, but analysts made the final call. Customer service bots answered questions, but humans handled anything complex. The AI was powerful, but passive, waiting for instructions, executing tasks, reporting results.

Then, starting in mid-2023, something shifted. AI systems began not just suggesting actions but taking them. The distinction seems subtle but proved transformative. When OpenAI introduced function calling for GPT-4 on June 13, 2023, language models gained the ability to not just generate text but to execute code, call APIs, and manipulate data. Anthropic followed with Claude 2.1’s beta tool use feature in November 2023, moving toward general availability with Claude 3 in early 2024. Suddenly, AI could act on the world, not just describe it.

By late 2024, the terminology had crystallized around “agents,” AI systems that pursue goals autonomously, make decisions without constant human oversight, and adapt their strategies based on outcomes. The shift wasn’t merely technical. It represented a fundamental reimagining of human-AI collaboration, raising questions about control, trust, and the appropriate boundaries of machine autonomy in commerce.

What Is an Agent?

The concept of an “agent” in computing long predated modern AI. In 1977, Carl Hewitt at MIT introduced the Actor Model, describing autonomous computational entities that could receive messages and make decisions. But those were programmatic agents, following strict rules. AI agents operate differently, they work toward goals, not instructions.

The distinction became clear through implementation. A traditional rules-based system for inventory management might follow 147 explicitly programmed decision points covering every anticipated contingency. An AI agent given the goal “maintain optimal inventory levels” develops its own approach: monitoring sales velocity, predicting demand, identifying supply chain constraints, and adjusting orders accordingly. The agent doesn’t follow a predetermined workflow, it constructs workflows based on the situation.

Function calling, introduced by OpenAI in June 2023, provided the technical foundation. Before this capability, language models could describe what actions to take but couldn’t execute them. Function calling allowed models to recognize when they needed to access external systems, databases, APIs, or code execution environments, then format proper calls to those systems. Anthropic’s implementation in Claude 2.1, announced in November 2023, took a similar approach, allowing models to use tools in combination and chain outputs.

The key characteristics of an agent, as they emerged through 2023 and 2024, include:

• Autonomy: Agents act without requesting permission for each step. Given a goal, they determine the sequence of actions needed to achieve it.
• Goal-orientation: Agents pursue outcomes, not procedures. The same agent might approach identical goals differently based on available resources or environmental constraints.
• Reactivity: Agents respond to changes in their environment, adjusting strategies when circumstances shift.
• Tool access: Agents can manipulate external systems, APIs, databases, file systems, to accomplish their objectives.

Learning and adaptation: Most agent systems simulate adaptation using memory tools or recorded results; true autonomous learning is still limited, though the mechanisms vary. Some systems update through explicit feedback loops, others through observation of results.

This differs fundamentally from chatbots, which respond to queries but rarely initiate action. It differs from APIs, which require explicit invocation for each function. And it differs from rules engines, which execute predefined logic regardless of context.

Multi-agent systems add complexity. Instead of a single agent pursuing a goal, multiple specialized agents collaborate, each with distinct capabilities and responsibilities. The systems communicate, sharing context and coordinating actions. By late 2024, frameworks emerged specifically designed for multi-agent orchestration, though implementation challenges remained significant.