I don't know what "build an agent" means
I'd rather think in terms of building blocks of LLM applications
I don't know what influencers mean when they say they “built 75 agents.” I'd rather think in terms of building blocks and work my way up:
1. A chat thread. Just me and an LLM going back and forth. To shape the conversation, I start the chat with text instructions (nerdspeak: “system prompt”).
2. A chat thread + a tool. Whether remote or local, a tool is just another member of the group chat. It can read my email, run a command, post to Slack, or just edit a text file. The LLM tags it, and it replies in the chat with the result.
3. A chat thread + a tool + a skill. A “skill” is the LLM pulling a book off the bookshelf: a set of text instructions it grabs to prompt itself in the chat thread.
4. All of the above + a file system. Now it’s Memento, where each chat thread can build upon the last. When the chat thread ends, the next one wakes up and finds the notes from before.
5. All of the above + a trigger. Somebody has to “hit enter” - usually it’s me - but it could be a scheduled task/heartbeat, an inbound WhatsApp, a webhook. Just another way of hitting enter instead of me.
6. “Subagent” is when your chat thread hits enter on… another chat thread. It’s like when you’re using ChatGPT to plan a vacation, and open a second ChatGPT tab to figure out what car to rent, then close the tab. This avoids cluttering the main thread’s context when it just needs the bottom line.
So, “building an agent” means giving an LLM a bunch of text files. Maybe a wrapped API or two. Occasionally, a file system. And someone/something that hits enter.
That’s the whole space. And heck, try to get away with as little as you can.
* * *
By a delightful fluke of history, the “agents” we run on our computers (e.g. inside Claude Cowork or Codex) are exactly the same as those running in the cloud (say, Notion AI, Zapier’s Shared Brain, or this Trello clone).1
How cool is that? If AI sucks at something in Claude Cowork after your best efforts, it's going to suck as a capital-A Agent. It’s the same building blocks. There’s no pixie dust.
On the plus side, the best way to reach the forefront of AI is to build your personal OS with Claude Cowork/Codex/Cursor etc. Precisely because you can directly touch everything in 1-6 in their UIs.
Imagine how strong your AI product intuition would be if you worked out of there daily. It’s my favorite way to learn. Best of all, you’ll see right through those “I built an agent” people.
PS If you’d like to do that in a hands-on, guided live workshop, see the links at the top of this post. Aman Khan (AI at Google, Spotify, Apple, Cruise) and I have two announcements:
Our live weekend sprint for executives and product leaders, Build AI Product Sense, starts in 9 days. $700 discount for readers of this newsletter.
We’re leading a free, live screenshare walkthrough of how to build your company AI OS (a great way to build your team’s AI product sense) this Friday.
See you there -Tal
I eeked in delight when I saw that Anthropic’s agents for financial services are the exact same text files whether you use them in Cowork, or run them in the cloud.