Agentic Artificial Intelligence

AI Agents

AI Agents are just LLMs that can take actions. The intersection between Language Models producing a reply based on an input, and programmatic workflows that can be triggered based on the reply, results in AI Agents
AI Agents cannot replace (for now), and can only augment.
Use AI Agents to automate the boring, rote and repetitive bits, and humans to verify and be creative to navigate around the complex bits.
E.g. In cybersecurity incidents, the AI agent pulls logs from different systems, join them via datetime, hostname, username, and draws more contextual information around them. As an analyst, this always needs to be done routinely, so we use an Agent to automate this
The human analyst then steps in and reasons around the story, and decides whether or not this story is malicious, suspicious or benign.

Architecture

Specialization

Try to make each agent specialized in one action
Don’t make an agent do multiple actions. This will make tuning an agent harder, as you may pollute the context of other actions

Linear

Actions are passed down linearly from one agent to another

Hierarchical

There is one central agent who delegates work to other agents
 

SPAR Framework

Every AI Agent
  • Sensing
    • The prompts or data fed into the LLM
    • Include contextual information, direct prompts, external information, web searches
    • Guardrails: Preventing unauthorized access of data.
  • Planning
    • Based on all the data, decide what is the best course of action
    • Guardrails: Preventing plans that are not human-friendly. Overoptimized for the objective, but is inhumane.
  • Action
    • Execute the action
    • Guardrails: Preventing unauthorized or destructive actions or
  • Reflection
    • Based on the output of the action, update the working memory of the agent that can be re-fed into the sensing stage

Milestones of Agents

Action

Reasoning

Memory