To consider ASI in realistic terms…
The Evolution of AI Agents
…it is helpful to consider how far we have come in few years’ time…
There were more simpler framework or technology prior to AI Agents only a few years ago…prompt engineering, templates, prompt chaining, RAG, etc…
If I look at my Medium timeline, it is littered with articles on these topics, and one can see based on the content how the Language Model scene has unfolded…
Before the rise of full-fledged AI Agents, which integrate planning, tool use, and autonomous workflows around large language models (LLMs), there were several simpler frameworks and techniques that laid the groundwork…
They acted as stepping stones, enhancing LLMs’ capabilities without the full complexity of agency.
The focused on improving accuracy, context handling and task decomposition in more structured but less autonomous ways.
The evolution roughly follows this sequence, building incrementally from basic prompting to more sophisticated systems.
It was as if there was a race, as we ere discovering new features and possibilities around Language Models, to build scaffolding, frameworks and SDKs to support new functionality.
I think this is why it remains hard to sell into enterprises and get to real ROI, because things are moving so fast…
Key Precursors to AI Agents
Their Progression
Here’s a breakdown of the main ones I mentioned in the recent past (prompt chaining, pipelines, and RAG), along with their timelines and how they evolved toward AI Agents.
Prompt Engineering (Early 2020s)
The simplest starting point after raw LLMs.
This involved crafting detailed wording and instructions to guide model outputs, like few-shot learning (providing examples in the prompt) or chain-of-thought (CoT) prompting, where the model is instructed to think step by step.
It was a manual way to elicit better reasoning without external tools.
By 2022, techniques like self-consistency (generating multiple responses and voting on the best) emerged, but it was still passive — relying on human ingenuity rather than automation.
Prompt Chaining (Mid-2022 onward)
Building on basic prompting, this technique sequences multiple prompts in a workflow, where the output of one becomes the input for the next.
It’s like breaking a complex task into steps (first summarise a text, then analyse it, then generate a response).
Early implementations were in tools like LangChain & Kore.ai (launched 2022), which allowed developers to chain prompts programmatically.
This added structure but required predefined sequences — no real adaptability or tool integration yet.
It was a key pattern in early “agentic workflows,” helping transition from static responses to multi-step processes.
Pipelines (2022–2023)
Pipelines refer to modular workflows combining LLMs with other components, often in a linear or branched setup.
For example, a pipeline might include preprocessing (data cleaning), an LLM call, post-processing (validation) and output formatting.
Frameworks like Hugging Face Transformers or early versions of Haystack enabled this, allowing integration of multiple models ( one for embedding, another for generation).
Pipelines introduced reusability and scalability but were rigid — designed for specific tasks like question-answering systems, without self-correction or external interaction.
Retrieval-Augmented Generation (RAG) (Late 2020–2023)
Introduced in a 2020 paper by Facebook AI (now Meta), RAG combines retrieval from external databases (vector search on documents) with LLM generation to reduce hallucinations and incorporate up-to-date knowledge.
It was a major leap, as it augmented LLMs with real-world data dynamically. Providing context at inference.
RAG was enabled by the discovery of in-context-learning. The principle that LLMs surrender or deprioritise the data they were trained on, in favour of information (contextual reference) supplied at inference.
Early RAG was simple (retrieve then generate), but by 2023, variants like MultiHop-RAG (handling multi-step queries) and advanced chunking/embedding techniques emerged.
This bridged to AI Agents by adding “memory” via retrieval, though it lacked planning or iteration.
These precursors evolved into AI agents around 2023–2024, when frameworks like Kore.ai, AutoGen, CrewAI and LangGraph started incorporating them into loops with tool calling, memory, and self-reflection.
For instance, Agentic RAG (emerging mid-2024, coined by LlamaIndex I believe) treats RAG as part of an agent’s toolkit, allowing iterative retrieval and reasoning.
AI Agents go further by adding proactivity (deciding when to retrieve or chain prompts autonomously) and multi-agent collaboration, turning static pipelines into dynamic system.
The Evolution of AI Agents
Considering the header image, it starts with basic language models focused on understanding and generation.
From here, layering on frameworks for execution, planning and tool use to create actual AI Agents.
Language Models as the Foundation
LLMs are the core, like GPT-series or Claude models, handling text generation and comprehension but lacking adaptability or interaction with the real world.
AI Agent Frameworks
This layer adds execution via tools (APIs for web browsing or code running) and planning (breaking down complex tasks into steps).
AI Agents
Here, systems become more integrated. They handle workflows through multi-agent collaboration or optimised sequences.
Evolving Agents
The exciting part — the data flywheel, feedback loops, fine-tuning, learning, and memory (or context) kick in.
AI Agents generate their own data for refinement (data flywheel), learn from interactions via loops like reinforcement learning and retain knowledge with memory systems like Mem0.
This creates a self-improving cycle, boosting adaptability.
Toward ASI (Artificial Super Intelligence)
The arrow points to super-intelligent systems that adapt at levels beyond human capability, with increasing agency through co-evolution of models and agents.
Nvidia coined the data flywheel — it’s a term popularised by Nvidia’s CEO Jensen Huang, who described it as a self-reinforcing loop where AI processes generate data to continuously improve models, especially in enterprise settings.
Nvidia has even built blueprints and microservices around it, like NIM Agent Blueprints, to help companies spin up these loops for agentic AI.
As of 2025, AI Agents have really taken off, moving from hype to practical deployment.
Reports show they’re becoming more autonomous, handling end-to-end tasks like project scoping, execution and adaptation without constant human input.
Chief Evangelist @ Kore.ai | I’m passionate about exploring the intersection of AI and language. Language Models, AI Agents, Agentic Apps, Dev Frameworks & Data-Driven Tools shaping tomorrow.
