Let's talk

The Agentic Shift

Key AI Milestones 2023 to 2026

While 2023 was defined by the explosion of generative AI, the period through 2026 marks a transition toward reasoning models and autonomous agents.

By Divyaprakash JhaUpdated 4 min readTopic: History of AI

Illustration: Key AI Milestones 2023 to 2026

Between 2023 and 2026, AI is shifting from simple chatbots that talk to users into autonomous agents that can reason through problems and perform multi-step tasks. This transition moves us from models that just generate text to systems that can actually use tools and complete complex workflows.

2023–2024: The Generative AI Explosion

The period starting in 2023 was defined by the sudden availability of high-powered generative AI for the general public. This era moved us away from the older eras of AI, such as the symbolic AI of the 1950s or the statistical machine learning of the 1990s, and into a world of massive foundation models.

In 2023, we saw the rise of open-weight models like Llama, which allowed developers to run powerful AI on their own terms. This was the same year GPT-4 was released, introducing multimodal capabilities—the ability to process more than just text.

By 2024, the focus shifted toward native multimodality. Models like GPT-4o and Gemini were designed to handle text, audio, and images all at once, allowing for real-time voice conversations that feel much more natural than previous interfaces.

We also saw the beginning of "reasoning at test time." With the release of OpenAI's o1 in late 2024, models began to spend more time "thinking" through problems step-by-step before they gave an answer. This improved how AI handles complex math, science, and coding tasks.

2024–2026: The Rise of AI Agents

As we move through 2024 and into 2026, the industry is moving from "chatting" to "doing." We are entering the era of AI agents—systems that don't just answer questions but actually complete jobs.

The Tools of Agency

  • Computer Use: In late 2024, models like Claude demonstrated the ability to actually operate a computer screen, moving the mouse and clicking buttons just like a human would.
  • The Model Context Protocol (MCP): Released by Anthropic in November 2024, MCP acts as a standard way to connect AI models to your specific data and software tools.
  • Agentic Coding: By 2025, coding agents have become standard tools for software teams. These agents can read an entire codebase, run commands, and make actual changes to the code to fix bugs.

For a small business owner in India, this might mean an AI agent that doesn't just write a customer service email, but actually logs into your accounting software, checks an invoice status, and updates a spreadsheet without you needing to prompt every single step.

Reasoning and Open-Weight Breakthroughs

A major turning point occurred in early 2025 with the release of DeepSeek-R1. This model proved that high-level reasoning—the ability to think through complex logic—could be achieved using open weights at a much lower cost than previously thought. This discovery reset industry expectations regarding how much money and compute is required to build frontier-level AI.

Because of these breakthroughs, the way we measure AI is changing. We are moving away from testing how "fluent" or "conversational" a model sounds. Instead, the new benchmark is how well a model can perform multi-step tasks, such as researching a deep topic, processing a batch of complex documents, or running a business workflow from start to finish.

The Infrastructure of Agency

For agents to work, they need "plumbing"—the standards that allow different AI systems to talk to each other and use the same tools. Two key developments are driving this:

  1. MCP (Model Context Protocol): This allows models to reach into your files and apps.
  2. A2A (Agent2Agent) Protocol: Announced by Google in April 2025, this protocol allows agents from different companies to find each other and work together on a single project.

At the same time, hardware is becoming more efficient. While massive models still require huge data centers, new "open MoE" (Mixture-of-Experts) models, such as Google's Gemma 4, are becoming highly capable while remaining small enough to run on a single workstation GPU. This means businesses can run sophisticated AI locally without needing to send all their private data to the cloud.

Global and National Milestones

The development of AI is not just happening in Silicon Valley; it is a global race involving significant national investment. One of the most notable examples is India's proactive approach to AI sovereignty.

In March 2024, the Indian government approved the IndiaAI Mission, a massive investment of ₹10,371.92 crore over five years. This funding is designed to provide shared GPU compute, support for homegrown foundation models, and training for the next generation of AI talent. This is critical for building models that understand Indian languages like Hindi and Marathi, which are essential for millions of small businesses in the country.

The global importance of this movement was highlighted in February 2026, when India hosted the AI Impact Summit in New Delhi. The summit concluded with the New Delhi Declaration, a significant international agreement endorsed by dozens of countries and organizations regarding the future of AI development.

Timeline Summary: 2023–2026

The shift from chatbots to agents can be summarized by the following key milestones:

Year Primary Focus Key Milestone / Technology
2023 Generative Chatbots GPT-4, Llama, and the rise of AI assistants.
2024 Multimodality & Reasoning GPT-4o, Gemini, and reasoning at test time (o1).
2025 Agentic Workflows Agent2Agent (A2A) protocol and agentic coding tools.
2026 Autonomous Agency & Standards Capable open models (Gemma 4) and the AI Impact Summit.

As we look toward the future, the trend is clear: AI is becoming less of a tool you talk to, and more of a digital worker that acts on your behalf.