Agents vs. Skills

A lot of the state of the world is trending towards using AI more efficiently and figuring out how LLMs interact with us, the people using them. Anthropic and Open AI are channeling millions upon millions of dollars into systems that squeeze more value out of the systems that they've built. Much of the changes that they're focusing on at this moment are context-related. Everyone is building systems that guess context more efficiently and try to understand the specific answers to the vague questions that they are asked. Fable pushes significantly harder towards guessing answers than any model I have seen before. Even Sonnet will spend 12,000 tokens trying to figure out the answer to a simple question. Gone are the days when assumptions are made, and instead now context is built automatically, and more accurately than ever before.

That said, context building remains the most valuable tool for getting what you want from LLMs. Understanding how the tools work and, more importantly, how to get what you want from them is always the right way to go about working with them. To that end, my mental model around agents and skills has become clearer lately.

Skills are for adding to context. Agents are for reducing it.

Let me explain what I mean. Most of us have experienced skills at this point and have some idea of what they are for. Any task you are going to repeat with an LLM should be considered a potential conversion to a skill. You already know the outcome you want. You've already done the deep dive on explaining it to an LLM, and you've gotten a result that you like well enough to write it down and make it something more repeatable. Creating a skill allows you to inject pre-created context more efficiently when it is needed. It allows you to tune a very non-deterministic system to be slightly more deterministic through the addition of context.

Agents, on the other hand, remove context that would be otherwise needed to fill in the blank. By limiting the scope of the potential paths that an LLM might take, agents allow context that exists to be more focused. For example, when you say "you are an incredible software engineer," The LLM can immediately get rid of every other profession that is not a software engineer. Saying, "you are an expert in financial modeling" allows the LLM to ignore all musical theory. LLMs are generalists. By nature, they contain information about every industry and every process. Agents allow those LLMs to ignore large swaths of their programming in exchange for focus where you want them to be.

What's interesting is that there is occasionally a temptation to make agents able to do many things, to add to their responsibilities and add to their skillsets. If we could just make an agent that is good enough, it would be able to handle whatever problem we threw at it. This is once again where Fable is leaning, and it is very expensive. The alternative is to make an army of agents, each of which has a tiny goal and focused skill set. Wrapping them up using an orchestrator agent is extremely effective. It allows your agents to execute and then stop before their context gets too bloated. It allows a retrieval of information and then execution separately from retrieval.

The smaller the job, the more focused the goal, the more likely the LLM is to produce the outcome you want.

Together, Skills and Agents work beautifully. If the skill is the weapon that is used to attack the problem, the agent is the wielder able to choose the right tool for the right job without distractions.

To take the metaphor a little bit further, delegation is a skill that I often associate with senior engineers. The ability to understand how to give work to somebody else and trust that they're going to get it done comes down to the ability to properly describe the outcome that is wanted. Taking the time to build the context and limit the chaff is the most valuable skill in prompt engineering and context wrangling today.

When LLMs make us feel that we are not producing enough and that we are capable of so much more, the return to our roots as guides and bolsters becomes more and more valuable.