LLM Agents vs. Workflows

Defining Agents

• Some consider anything beyond a single LLM call an "agent."
• A true agent involves the LLM autonomously deciding how many times to run a process.
• Agents loop until a resolution is found, such as in customer support or code iteration.
• The LLM picks its own path and actions, instead of following a predefined path.
• Workflows, in contrast, follow a fixed number of steps and are considered "on rails."

Evolution of Agents and Workflows

• The distinction between agents and workflows evolved as models became more sophisticated.
• Initially, systems used a single LLM, progressing to multiple LLMs orchestrated by themselves.
• Workflows are pre-orchestrated by code, while agents are simpler but more complex in different ways.
• As models and tools improve, agents are becoming more prevalent and capable.

Coding Differences

• Workflow prompts involve a straight line of steps, from prompt A to B to C.
• Each prompt in a workflow is specific, transforming one input into another output.
• Agent prompts are more open-ended, providing the model with tools like web search or code editing.
• The agent continues using these tools until it finds an answer.

Designing Effective Agents

• One must empathize with the model by understanding its context and knowledge limitations.
• Instructions, tool descriptions, and the environment must be clear to the model.

Tool Use and Prompt Engineering

• Developers should not neglect prompt engineering for tool descriptions.
• Tool descriptions should be created with a classical programming system in mind

Motivations for Defining Agents

• Need for clear definitions to facilitate conversations with customers.
• Models are now capable of handling many agentic workflows.
• Guide people about how to do agents but also where agents are appropriate

Underhyped Aspects of Agents

• Automation of tasks that save even a small amount of time can have a big impact.
• Automating small tasks enables scaling those tasks to a much greater extent.
• Calibrating agents to where they are needed is difficult.
• Intersection of valuable, complex tasks is a sweet spot.

Agentic Coding

• Agentic search: Precision can be traded off for recall, getting more information than needed.
• Results on SWE-bench (coding benchmark) have increased from very low to over 50%.
• Greatest block to coding performance is verification, real world unit tests are not perfect

Multi-Agent Environments

• Potentially, multi-agent environments involve multiple agents interacting and coordinating.
• Challenges include determining how agents interact and whether this is better than a single agent.
• Barry built an environment where multiple Claude models play Werewolf together.
• Werewolf demonstrates interesting interactions and emergent behaviors among agents.

Future of Agents in 2025

• Widespread business adoption of agents is expected, automating repetitive tasks.
• Agents will enable scaling up tasks that were previously too expensive.
• Examples include coding agents updating documentation with every pull request.
• Agents for consumers may be overhyped due to difficulties in specifying preferences and high verification costs.

Advice for Developers

• Create processes to measure results, ensuring you know if your system is working or not.
• Start as simply as possible and add complexity, this will remain the case even as the models improve.
• Those excited by improvements to the models will see success, but those afraid because their moat will disappear are building the wrong thing.

Explore the distinction between LLM agents and workflows, focusing on autonomous decision-making versus predefined steps. Understand the evolution of these systems and their coding differences. Learn how agents are becoming more capable with improving models and tools.