This is exactly what I was looking for, to give it to one of my mentees as agents 101. Thanks much for writing and covering so much together in one article. This is very very helpful.
This was a great read. The distinction you draw between one-shot model usage and agents really clicks when you frame it as software shifting from a reactive tool to something that can actually pursue goals over time.
What resonated most for me is how agents change the economics of work. A lot of valuable tasks were never automated not because they were impossible, but because they were too messy, too exception-heavy, or too small to justify traditional engineering effort. Agents feel like they finally make that long tail approachable.
I also appreciate the emphasis on context, memory, and tool use as first-class concerns. Once an agent can plan, act, observe results, and adapt, the mental model stops being “better autocomplete” and starts looking much closer to how teams actually operate.
The interoperability angle is especially important. Without shared protocols, we risk fragmenting the ecosystem before it really matures. MCP and A2A feel like the kind of boring but necessary foundations that determine whether this becomes infrastructure or just a collection of clever demos.
Overall, this piece does a great job balancing excitement with realism. Agents clearly unlock new leverage, but they also shift responsibility toward humans defining goals, constraints, and quality bars. That tradeoff feels like the real story here.
Very nice article highlighting everything about AI Agents. Thanks Addy Osmani for this wonderful writeup.
It's my pleasure. Thanks for the kind words and for reading!
Did you use Claude or GPT-4o to help write it? Curious on your writing process with LLMs.
So what’s the role of the human engineer 👷♀️ in this agentic era?
This is exactly what I was looking for, to give it to one of my mentees as agents 101. Thanks much for writing and covering so much together in one article. This is very very helpful.
Beautifully penned.
More power to you.
you are an absolute gem Addy!
bookmarked, need to revisit this when I have more time to dig in
This was a great read. The distinction you draw between one-shot model usage and agents really clicks when you frame it as software shifting from a reactive tool to something that can actually pursue goals over time.
What resonated most for me is how agents change the economics of work. A lot of valuable tasks were never automated not because they were impossible, but because they were too messy, too exception-heavy, or too small to justify traditional engineering effort. Agents feel like they finally make that long tail approachable.
I also appreciate the emphasis on context, memory, and tool use as first-class concerns. Once an agent can plan, act, observe results, and adapt, the mental model stops being “better autocomplete” and starts looking much closer to how teams actually operate.
The interoperability angle is especially important. Without shared protocols, we risk fragmenting the ecosystem before it really matures. MCP and A2A feel like the kind of boring but necessary foundations that determine whether this becomes infrastructure or just a collection of clever demos.
Overall, this piece does a great job balancing excitement with realism. Agents clearly unlock new leverage, but they also shift responsibility toward humans defining goals, constraints, and quality bars. That tradeoff feels like the real story here.