The Promise of Agentic AI

Last week's CloudHustle featured a mention of Claude Code, a new agentic coding assistant for the command line from Anthropic. 

I spent the weekend using Claude Code to develop a small software project, and the experience got me thinking. 

LLMs are excellent at generating content, performing summarization, and other content generation use cases. This should clearly come as no surprise, given that it's called generative AI. But my experience with Claude Code has really reinforced the promise for agents to take action, with LLMs focused on orchestration.

Claude Code Capabilities

Claude Code comes with a variety of different "tools" that are effectively agents. Examples of such tools include:

• BashTool, which can execute shell commands
• GlobTool, GrepTool, and LSTool, which can search for files, search for file content, and walk your project's source tree
• FileReadTool allows the LLM to request content from an existing file, very much in the same way as Retrieval Augmented Generation (RAG) FileEditTool and FileWriteTool can make changes to existing content in the project, and also create new files and directories as necessary

While BashTool has a very short description, it brings a massive amount of power and capability to Claude Code, taking advantage of the large ecosystem of command line tools, especially on platforms like Linux and macOS, which offer many common tools due to their shared UNIX-ish lineage. This means that the BashTool agent is capable of interacting with git, curl, compilers, linters, etc.

Why is this useful?

Well, let me provide some examples.

If you're a programmer like me, you know just how complicated git is. While an extremely popular source control platform, git is notoriously obtuse, with many options, inconsistencies, and warts. 

Whenever I have to go beyond the day-to-day basics with git, I have to spend time refreshing my memory. Now, these days I often turn to LLMs to help me craft complex CLI commands, but that is very different from having agents execute those commands on my behalf. Over the weekend, I realized that I had accidentally committed an API key to my project's git repository. 

Thankfully, it's possible to rewrite history, but it's kind of a painful process. This time, I just told Claude Code to do it for me. And it did. Voila!

Use Cases for the Average Joe

Developer use cases are interesting, but what about use cases for your average consumer?

Apple has phones in the hands of millions, and iOS has a robust and unparalleled ecosystem of apps. For many people, their phone may be their most used computer, if not their only computer. 

This puts Apple in a unique position, allowing developers to implement "App Intents", which provide a foundation for future agentic behavior baked right into the OS. Agent-driven actions within apps could be extremely impactful. Apple's execution in GenAI has been lousy, drawing harsh criticism of even the most pro-Apple pundits, but there have long been rumors that Apple is aiming to aggressively pursue this agentic strategy, which seem highly plausible given their other behavior. While Apple may not get there soon, they will eventually, but their timing leaves the door open for competitors.

Agents are the Future

Putting this all together, the lion's share of GenAI workloads today are heavily focused on information, asking ChatGPT, Claude, Gemini or other such AI assistants to answer questions conversationally.

But, ultimately, that's only a small piece of the puzzle. 

While providing information is certainly useful, I am of the opinion that the long-term arc of GenAI is heavily bent toward agentic AI, as our assistants evolve from providers of information to action-oriented collaborators. 

Food for thought!