Since 2023 I’ve logged more than 5,000 hours of hands-on work with generative AI: chat assistants first, then coding assistants, and now agents that plan, use tools and work for long stretches on their own. Before that I spent more than two decades building software, starting in 1999. That background matters, because the biggest lesson of the last three years is that working agentically is a management skill more than a technical one.
Here’s what I’ve learned about getting real work done with agents, not demos.
From prompts to missions
A prompt asks for an output. A mission asks for an outcome. “Write a function that parses this file” is a prompt. “Make the nightly import reliable: find why it fails, fix it, add a test that would have caught it and tell me what you changed” is a mission.
Agents do their best work on missions, as long as the mission is well defined. A good mission has four parts:
- The goal: what should be true when the work is done
- The context: what the agent needs to know that it can’t discover on its own
- The boundaries: what it must not touch, spend or change without asking
- The definition of done: how you and the agent will both know it worked
I’ve written before about the automation of automation, and missions are where that idea becomes practical. Writing a clear mission takes ten minutes. It routinely saves hours.
Assign work like a manager, not a typist
The shift that made the biggest difference for me was to stop treating the agent as a faster keyboard and start treating it as a capable, very fast, sometimes overconfident new team member.
That means:
- giving it whole tasks, not fragments
- letting it choose the approach for routine work
- breaking big work into pieces that can each be checked on their own
- running independent pieces in parallel when the tools allow it
- staying available for questions instead of hovering over every step
It also means matching the task to the agent. Some work suits a long-running autonomous agent. Some suits a quick back-and-forth. Some still suits doing it yourself. Knowing the difference is part of the skill.
Keep a human in the loop where it counts
Oversight doesn’t mean approving every keystroke. It means putting human judgment at the points where mistakes are expensive or hard to undo:
- anything that touches production systems, customer data or money
- anything sent to another person on your behalf
- decisions about scope, priority and trade-offs
- deleting or overwriting things
Everywhere else, let the agent work. The goal is the least oversight that still keeps you safe, not the most oversight you can manage. Too much and you’ve rebuilt the job by hand. Too little and you’ll find out the hard way.
Validate results, don’t trust them
Agents report success with the same confidence whether they succeeded or not. Early on, I lost more time to “done” work that wasn’t done than to any other problem.
What fixed it:
- Make verification part of the mission. Ask for evidence, not assurance: test results, links to check, before-and-after numbers.
- Check the work, not the summary. Read the diff, open the page, run the query yourself.
- Use tests and checklists the agent can’t talk its way past. Automated tests, validators and preflight scripts turn “trust me” into “here’s the result.”
- Stop after repeated failures. If an agent fails at the same thing two or three times, stop and rethink the approach together instead of letting it loop.
Give agents memory, carefully
Agents forget between sessions unless you give them a way to remember: project instruction files, notes, logs and change records. A short, well-kept instructions file that says how your team works is one of the highest-value things you can write.
Memory needs maintenance. Stale or hidden instructions cause confusing behaviour, as I found when hidden auto-memories broke several of my projects. Keep memory visible, short and current.
Measure whether it helped
It’s easy to feel productive with agents. It’s harder to prove it. Pick a few measures before you start and check them honestly:
- time from request to verified result
- how often work has to be redone
- defects that reach users
- how much of your own time went to reviewing versus doing
Some tasks will get dramatically faster. Some won’t improve at all. A few will get worse. Knowing which is which is how you decide where to use agents next.
What this means for businesses
The same lessons apply at company scale. Teams that get value from AI agents define clear missions, keep humans responsible for judgment, build verification into the workflow and measure outcomes. Teams that skip those steps end up with impressive demos and little to show in the results. I look at where enterprise adoption is heading in Enterprise AI Agent Adoption: Bottom-Up Research and Forecasting, 2026–2030, and at a concrete example of handing work between coding agents in Stop Copy-Pasting AI Coding Context.
Working agentically isn’t about replacing people. It’s about people doing more of the work only people can do. That’s the practical core of Human + AI, and it’s what we help clients put into practice at IT Strategists.