Are we building AI agents to do the right work, or just automating the old way of working?
I've been following the AI agent launches here, and one question keeps coming to mind.
We're building agents that write code, take meeting notes, conduct research, and execute workflows. In manufacturing, robots are already performing physical work.
But are we rethinking how organizations operate, or simply teaching AI to perform the same processes faster?
After more than a decade in IT project management, I've seen how much time organizations spend in meetings, preparing reports, coordinating work, and waiting for decisions.
What if some of those processes don't need to exist anymore?
For those building AI agents, where do you start?
Do you automate an existing workflow, or first question whether that workflow is necessary?
And how do you technically determine whether your agent accomplished the right outcome, rather than just completed its assigned tasks?
I'd love to hear how other builders are approaching this.
Replies
In my own setup the agents removed steps more often than they sped them up. There is no status report anymore, because the commit log is the report. There are no handoff notes, because every session starts by reading the project's notes file. On your last question, the check that works best for me is verifying the outcome from outside the agent: after a change lands, the agent has to run the app on an emulator and look at the actual screen, not just tell me the build passed. "Task completed" and "the thing works" are different claims, and only the second one is worth anything. Which processes did you see disappear first in the teams you've worked with?
For me it's well past 90% now. Email, customer support, documents, programming, all of it runs through AI agents.
On knowing whether an agent got the right outcome, I try not to let the LLM decide much in the first place. Anything that can be a script becomes a script. When an LLM does have to make a call, a different one checks it: Gemini checks Claude, and for the bigger ones Codex and Grok too. And for code I trust E2E tests a lot more than reading the code.