AI models didn't always know how to use tools, and the journey from that limitation to today's autonomous agents unfolded across dozens of research breakthroughs and vendor announcements between 2022 and 2025. A newly published timeline traces this evolution through primary sources, showing how large language models (LLMs) progressed from reasoning-only systems to agents capable of calling APIs, controlling computers, and coordinating with other agents through standardized protocols.

What Exactly Is Tool Use, and Why Does It Matter for AI?

Tool use, also called function calling, is the ability of an AI model to invoke external functions, APIs, or graphical interfaces by emitting a structured request that an application executes on the model's behalf. Without this capability, AI models could only generate text. With it, they can take action in the real world, retrieve live data, perform calculations, and coordinate with other systems. This shift transformed AI from a text-generation engine into an autonomous agent capable of solving multi-step problems.

The timeline begins in October 2022 with the ReAct paper, which introduced the concept of interleaved reasoning and acting. ReAct showed that prompting a model to alternate between reasoning traces and actions let it call external sources such as a Wikipedia API, establishing an early template for tool-augmented and agentic LLMs. Four months later, Meta AI published Toolformer, demonstrating that language models could teach themselves to call APIs in a self-supervised way, deciding which API to call, when to call it, and with what arguments.

When Did Commercial AI Platforms First Adopt Tool Use?

The first widely used productized "tools for a language model" arrived in March 2023, when OpenAI announced ChatGPT plugins. The alpha release let ChatGPT use first-party tools, including a web browser and a code interpreter, alongside third-party plugins and an open-sourced retrieval plugin. Three months later, in June 2023, OpenAI introduced function calling in its API, allowing developers to describe functions to GPT-4 and GPT-3.5 Turbo, with the model returning a JSON object containing arguments to call them. This became the first mainstream function calling API primitive.

Anthropic followed a different path. The company added tool use as a beta feature with Claude 2.1 in November 2023, introducing a 200,000-token context window, system prompts, and an initial beta of tool use on the Anthropic API. After entering public beta across the Claude 3 family in April 2024, tool use reached general availability on Claude 3 Opus, Sonnet, and Haiku across the Claude API, Amazon Bedrock, and Google Cloud Vertex AI by May 2024. That same month, AWS launched the Amazon Bedrock Converse API with standardized tool use, providing a single consistent interface across foundation models.

OpenAI continued advancing the capability in August 2024 by introducing Structured Outputs, which guarantees that model output conforms to a developer-supplied JSON Schema, building on the earlier JSON mode and enabling strict function calling.

How Did Tool Use Expand Beyond APIs to Graphical Interfaces?

The next major evolution came in October 2024, when Anthropic released computer use as a public beta. Announced alongside an upgraded Claude 3.5 Sonnet and Claude 3.5 Haiku, computer use let Claude perceive a screen and control a mouse and keyboard, extending tool use from APIs to graphical interfaces. This capability allows AI agents to interact with software the same way humans do, opening new possibilities for automation across legacy systems that lack APIs.

What Are the Two Open Protocols Now Structuring the Agent Ecosystem?

As tool use matured, the industry recognized the need for open standards to coordinate how agents access tools and how agents communicate with each other. Two protocols emerged as foundational:

• Model Context Protocol (MCP): An open standard that lets AI applications connect to external data and tools through a uniform interface, standardizing the vertical relationship between a model and its tools and data sources.
• Agent2Agent (A2A) Protocol: An open standard that governs how agents discover and delegate work to each other, standardizing the horizontal, agent-to-agent relationship.

Anthropic announced and open-sourced the Model Context Protocol in November 2024, releasing it with Python and TypeScript SDKs and reference servers for Google Drive, Slack, GitHub, Git, and Postgres, among others. The initial MCP specification, dated November 5, 2024, defined the client-server model, JSON-RPC messaging, and the tools, resources, and prompts primitives.

Adoption accelerated rapidly. Microsoft introduced MCP in Copilot Studio in March 2025, letting makers connect agents to existing knowledge servers, APIs, and other MCP-enabled data sources. More significantly, OpenAI adopted MCP starting with its Agents SDK on March 26, 2025, marking the first adoption of the standard by another frontier AI lab. The MCP specification was revised on March 26, 2025, adding an OAuth 2.1-based authorization framework and replacing the HTTP plus Server-Sent Events transport with a more flexible Streamable HTTP transport.

How to Understand the Practical Implications of These Protocols for Developers

The emergence of standardized protocols has concrete implications for how developers build and deploy AI agents. Consider these practical dimensions:

• Reduced Vendor Lock-In: Open protocols like MCP and A2A allow developers to build agents that work across multiple AI platforms and models, rather than being tied to a single vendor's proprietary implementation of tool use.
• Faster Integration: Reference servers and SDKs for common data sources (Google Drive, Slack, GitHub, Postgres) mean developers can connect agents to business systems without building custom integrations from scratch.
• Interoperability at Scale: As more platforms adopt MCP and A2A, agents can discover and coordinate with each other automatically, enabling complex multi-agent workflows without manual orchestration.
• Standardized Security: OAuth 2.1-based authorization in MCP provides a consistent framework for permission management across agents and tools, addressing a critical concern for enterprise deployments.

The timeline also reveals a broader pattern: tool use capabilities have moved from experimental research to production-grade features in less than three years. What began as a research question in 2022 has become the foundation of a new category of software. The shift from proprietary implementations (OpenAI function calling, Anthropic tool use) to open protocols (MCP, A2A) suggests the industry is settling on shared standards for how agents interact with tools and each other.

For enterprises and developers, this timeline matters because it shows where the technology is headed. The convergence on open protocols signals that tool use and agent coordination are no longer experimental features but core infrastructure for AI applications. Organizations building AI agents today should expect that the tools they choose will need to support these emerging standards, and that the agents they deploy will eventually need to interoperate with agents built on competing platforms.