Agents and Agent Platforms

AI agents are programs where LLM outputs control the workflow.

An AI solution is one that involves any of these:

1. Multiple LLM calls
2. LLM with the ability to use tools
3. An environment where LLMs interact
4. A planner to coordinate activities
5. Autonomy

Agentic Systems

Anthropic distinguishes two types:

1. Workflows: Systems where LLMs and tools are orchestrated (planned, organized, and coordinated very carefully—often from behind the scenes—to achieve a highly specific and desired outcome) through predefined code paths. It follows a predefined, step-by-step sequence of operations (like a recipe).
2. Agents: Systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.

Workflow

They have 5 design patterns:

1. Prompt Chaining: Decompose into fixed small tasks. (Input goes from one LLM to the next in a linear manner).
2. Routing: Direct an input into a specialized sub-task, ensuring separation of concerns. (Input goes to one LLM, and then the LLM decides which model to select next).
3. Parallelization: Breaking down tasks and running multiple tasks concurrently. (Input goes to a program, the code decides to send it to multiple LLMs for parallelization, and then an aggregator aggregates the output).
4. Orchestrator-Worker: Complex tasks are broken down dynamically by an LLM and combined by a synthesizer.
5. Evaluator-Optimizer: LLM output is validated by another LLM.

Agents

1. Open-ended
2. Feedback loop
3. No fixed path (dynamic, no fixed steps)

• Note: They are less predictable, having no predictable paths, output, or cost. Thus, use monitors and guardrails.

A human feeds an input to the LLM, and then the LLM interacts with the environment and gets feedback from the environment.