AI Agent Definition

#

This serves as both a historical overview and a implementation definition of what constitutes an AI agent and its core components. It is a loop regardless of what it morphed into.

Andrew Ng (March 2024)

#

Andrew Ng started talking about agentic workflows basically having an army of “AI agents” (LLms) that could retrieve information and help with research. Around March 2024, her wrote a linkedin post about AI agents

Implementation

#

The whole thing boils down to giving LLMs (AI agents) different personalities and roles through system prompts. It’s basically getting AI to work more like humans do, bouncing ideas off each other, revising, and specializing instead of trying to do everything at once. Here’s how it works:

• Persona injection: Each LLM agent gets a specific role and expertise area. You ask it to roleplay for example as a developer, a tester, a researcher etc.
• Loop: Each of the output of the LLM is given as input for another LLM critique each other’s work and iterate. Back then the context windows of most LLms are not that huge, so you could not do much.
• Completion criteria: Define rules for when they’re done.
• Multi-agent collaboration: Different agents handle different parts of the task.

Satya Nadella keynote defitnition of AI (Microsoft Ignite on November 19, 2024)

#

By late 2024, the defition of AI agents has transformed yet again. This time it was laid out by microsoft Microsoft Ignite 2024 Keynote - November 19, 2024 . Removing the marketing buzz words and so, It is centered around creating an “agentic world” where AI acts on your behalf (similar to Andrew Ng). Nadella outlined three AI capabilities: universal interface, advanced reasoning/planning, and long-term memory. The universal interface allows AI to accept and process various types of input, including speech, images, and videos. AI models can now retain rich context over extended periods (i.e. a RAG).

Implementation

#

I think an AI agent now in late 2024 - early 2025 has become the following. It evolved from personalities to this. But it is still a loop. This is how I would implement it.

• Agents are multimodal but mostly tend to be an LLM, can take information from different models providing capabilities and sources and work with it.
• Multi agent framework would be many versions of the single agent pieced together in series or in parallel.
• There will be an orchestration agent that handles all the responses.
• A single AI agent is comprised of:
  • A start and a single/multiple end points.
  • A loop
  • A performance criteria.
  • Memory or a series of preferences (RAG).
  • A decent enough LLM with large context windows.
  • Access to different tools such as a terminal, code editor, runtime environment, access to data, write and execute tests.
  • A series of steps to capture thinking and working through a problem. Examples include planning.
  • A group of prompts that it can pick and choose depending upon the situation.

Thoughts (Updated in August 2025)

#

There are a lot of tools and companies now. But a clear definition is still lacking.

Context size is everything

#

It really does not matter which of the definitions you use. The LLM context size decides whether your agentic framework will fail or succeed. The larger the context size is, the better the agent should perform. But in practice, the worse an LLM performs, and it has not changed much over the last year. Despite some of the best models like Claude now supporting a million tokens, in practice they get worse when going above 100k tokens.

Isolation of subagents

#

Each subagent needs to run in complete isolation to prevent context corruption. It also saves tokens in the long run and prevents unwanted usage.

Costs.

#

For many of these frameworks, it is very expensive to run blind agent queries.