· 1 min read
Managing Agents Works Like Managing People
Nothing about AI changed the fundamentals of software delivery—it just made clear specs, bounded work, and real validation impossible to fake.
Uber made waves burning a year's AI budget in months. Every company on earth is debating the role of token budgets, with extremes on both ends. Undisciplined experimentation and exploration with AI tools yields results about as good as a generic, no-context, one-shot prompt—obvious AI slop.
I shied away from early prompt engineering techniques. "You are an expert data architect. You only give right answers. You will not hallucinate." I still roll my eyes at those, though I've come to appreciate the value of well-structured prompt harnesses.
Managing agents and agentic work starts with appreciating that prompting and conversational practices are about as effective as they would be with people. You need to be specific. You need to give context. You need to be clear about what you want and expect. You also need layers of management—at a minimum, a tech lead and manager mindset to design, validate, and qualify the work.
Multi-model is non-negotiable. Breaking the work into concrete, attainable chunks (this isn't new) is also table stakes. No one-shots.
Ultimately, you need to be systematic about the way you approach development. Nothing about AI changed the rules here—it just reinforces the necessity of clear, clean planning, specifying, and validating.