Why It's So Hard To Make Useful Agents
Coming off Labour Day, there's a work-related theme that translates well to how we should think about working with AI agents:
At the turn of the last century, an American engineer named Frederick Winslow Taylor was revolutionizing the way work was done in factories. In his theory of Scientific Management, a process engineer would break done the various tasks to such a degree that they could then be performed by relatively low-skilled labour. He was so influential that this way of looking at work came to be know as Taylorism.
Fast forward to the mid-century, and a whole new class of work has emerged: Knowledge work, where people work on reports, or ad campaigns, or editing. In short, using their brain to add value to information. It's easy to see that Taylorism doesn't work well here. You cannot break down "craft a compelling billboard ad" into mechanistic steps that anyone could perform. And this is exactly what Peter Drucker—who coined the term knowledge work—pointed out: You need to give a knowledge worker a tight brief about what you want them to do, but give them autonomy over how they do it.
His ideas have informed our collective approach to knowledge work ever since, summarized in ideas like: "Hire the best people, then get out of their way."
And that's where the rub is with AI agents. They are decidedly not "the best people". If you don't give any specifics whatsoever on how you want them to accomplish something, all bets are off. They lack the good judgment and common sense of a genuine person. To work around that, we're encouraged to provide a lot of context to the model: Do this. Don't do that. ALWAYS do this. REALLY PLEASE NEVER DO THAT.
This is applying Taylorism to knowledge work, and it won't be fruitful for truly creative tasks.
Some of what gets lumped together with knowledge work is pretty mechanistic:
Scan this report for these numbers and put them into that spreadsheet
Scan this novella-length email from my school's administration for concrete tasks
Those are great fits for AI.
Others would benefit from some judgment, but are also sufficiently low-stakes that the efficiency gained is worth it:
Take this press release (one of many dozens we churn out each day) and adjust style and length to be used as a tweet.
Those can be handled by AI provided it gets a bit of extra context or embedded into the right harness.
But anything beyond that and you need that whole-human touch. Things like,
Prepare a pitch deck for my startup idea so I can get funded
Craft an outreach email to this former corporate client of ours
By the time you've explained to the AI exactly how it should do it, you'd have done a better job yourself and you wouldn't have cheated yourself out of the learning and growth that comes out of wrestling with such profound tasks.
By all means, use AI for the two lower tiers of knowledge work, and keep it away from the highest tier.
