Dave Ebbelaar

AI Consultant & Educator | Teaching developers how to build production AI systems and go independent | 240K+ on YouTube

[Dave Ebbelaar @daveebbelaar]

Great question. Short answer: yes, for client work we still prioritize workflows. What's changed is that at edge nodes, we introduce more and more tool calls. Up to five or so. This way you keep mostly deterministic flows but avoid building Frankenstein DAGs and just let the LLM figure it out in a loop with tools at the edges.

I'm also super excited about things like OpenClaw and Anthropic's Agent SDK. I am heavily testing it for my own automations and systems. But not for client work yet. It's not quite ready for that. Getting there though.

The future is agentic. We're just not fully there for production client work today.

[Dave Ebbelaar @daveebbelaar]

ChatGPT has been getting worse. Everyone's saying it. And they're probably right.

Meanwhile, the people actually building with AI are watching something completely different happen.

Most people interact with AI through a chat window. Type a question, get an answer. When a new model drops, they test it the same way.

The improvements feel marginal. In a chat interface, they are.

But that's not where the action is.

Over the past week, I ripped out every automation in my business and rebuilt them from scratch with the Claude Agent SDK.

Not simple if-this-then-that flows.

Agents that take a set of instructions, connect to an API, and figure out the execution path themselves.

The Agent SDK gives you the same tools, agent loop, and context management that power Claude Code, programmable in Python and TypeScript.

Claude Code started as a coding tool. It's already way beyond that. Give it instructions and access to an API, and it goes off and does the work.

When you can spawn these agent sessions programmatically, things get powerful fast.

Most people just pay attention to the models, but the real shift is happening because of three things converging:

- The models are more capable.
- The agent harnesses around them are maturing.
- And more services are shipping AI-first APIs that agents can actually use.

Those three combined are what's causing the acceleration. Almost everything you do behind a computer can already be automated.

Things are moving a lot faster than most realize.

[Dave Ebbelaar @daveebbelaar]

Anyone else find themselves constantly working on 2-3 projects at the same time with Claude Code?

Not sure if it's productive but I can't help squeezing in voice-transcribed prompts while waiting for outputs 🤷‍♂️

[Dave Ebbelaar @daveebbelaar]

That's a wrap!

If you enjoyed this thread:

1. Follow me @daveebbelaar for more of these
2. RT the tweet below to share this thread with your audience

[Dave Ebbelaar @daveebbelaar]

Making an LLM API call is the most expensive and most dangerous operation in modern software development.

While incredibly powerful, you want to avoid it at all costs and only use it when absolutely necessary.

Here's how to build AI agents that actually work: 🧵

[Dave Ebbelaar @daveebbelaar]

Making an LLM API call is the most expensive and most dangerous operation in modern software development.

While incredibly powerful, you want to avoid it at all costs and only use it when absolutely necessary.

Here's how to build AI agents that actually work: 🧵

[Dave Ebbelaar @daveebbelaar]

P.S. I explain all of this in detail in my latest video: https://t.co/LCsSrtITy9

[Dave Ebbelaar @daveebbelaar]

Here's my take on building effective AI Agents (without the hype)

https://t.co/jKBwBQYp7m

[Dave Ebbelaar @daveebbelaar]

I think PydanticAI is finally an agent framework that I actually like.

https://t.co/AwItFNFT2E

[jason @jxnlco]

This is a great video going over some of the validation logic in instructor that can't be done in structured outputs alone from openai ;)

https://t.co/UJAi1rVfLQ

[Dave Ebbelaar @daveebbelaar]

@jxnlco Thanks for the shoutout! 🙏🏻