My research agent, Marie, scored a D on its nightly report card for failing to finish assigned tasks.
While a D would conclude the agent was the problem, the real problem was a rule I had written: any task running past ten minutes should time out. The work was heavy and the agent spent many tokens on it, only to hit a wall midstream. It earned the D the way plenty of human team members earn theirs in performance review—by working under the constraints of decisions they didn’t create, and taking the fall for them.
My independent grader, Wanda, found it. Wanda’s only job is to watch for process failures. Wanda’s has no horse in the race to defend a bad design decision. Every night it sends me a grade for every agent that worked the prior day: error counts, frozen tasks, blocks, resource use, and completion times, along with its reasoning and a recommendation for what to change. I check its recommendations against the data and apply the correct fixes. Rather than a flawed performance review process that could penalize an undeserving team member, Wanda is very honest to share with measurable proof who holds the blame. In this case, it was I.
The cost bad managers never paid
After many years of working alongside managers, I have watched organizations run around poor ones. Capable team members routed around the problem despite the poor management they received. They found peers who bridged the managers’ gaps, made consequential calls themselves, and delivered quality outcomes. The bill for the manager arrived later as attrition. They don’t correct their ways, but only hire a replacement, and it is Groundhog Day again.
An agent has no peers to fill them in productively. It operates off the instructions, context, and boundaries it receives, and the quality their management shows up very quickly.
While model intelligence continues to improve, they remain predictive, which means they complete patterns, and they do not read minds. Leave the oven temperature off a recipe. While an experienced baker fills the gap from their years in the kitchen, a model fills it with whatever temperature seems most likely to it. People hear “intelligent” and assume human-style inference. In my experience, the connections a person makes without thinking are the same ones a model misses.
The same foundation, a faster clock
Leading agents and leading people overlap. In a pre-registered Harvard Kennedy School experiment, how well participants led AI agents predicted how well they led human teams, with a correlation of 0.81. The takeaway for CHROs: your strongest managers hold a head start, and your weakest managers’ shortcomings will be illuminated in a new and damning way.
With AI agents, management cadence and evidence change on a much faster clip, and ignoring problems for too long can compound exponentially and quickly. In my build, feedback arrives nightly, by design, and every adjustment rests on measured results using a rubric I built. A change is validated on a test agent before it reaches production,
then retested.
My agent management falls into three kinds:
1. Direction means writing down what the agent is, what good looks like, and the context about the person it serves.
2. Structure means changing workflows, handoffs, and clarity of agent ownership. I gave too much to too few agents across too many topics, results suffered, and one agent sharing two workloads created a backlog. So now, one builds the user product, and another builds the database. Releasing barriers means removing obstacles I had built without realizing it. The timeout and the vague handoff barriers were also eliminated in this phase of management.
3. Investment must be significant upfront and throughout. Think: go slow to go fast. In the early months, I spent more time coaching agents than building product, and some weeks it took at least 20 hours. Today every agent that started at a D scores an A. I spend more time on production builds and less on corrections. The grader still runs every night, so the work has shifted from fixing to monitoring and iterating.
What happens when leaders skip this discipline with both humans and agents?
Gallup’s 2026 data shows managers decide whether AI delivers value. Team members whose managers actively support AI use are far more likely to say it improves how work gets done: 33% versus 9%. In the same research, half of the CHROs in Gallup’s survey say they lack confidence in managers’ ability to guide team members’ AI use, and 57% are providing AI training for managers, which Gallup notes is unlikely to be enough on its own. That research covers managers guiding people’s use of AI. Running agents is a further step, and in my experience few programs have started to teach it.
Gartner forecasts that more than 40% of agentic AI projects will becancelled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Someone must own each of those. In a company running agents, that someone is a manager.
What this requires of CHROs
This is a new job requirement for managers, and the CHRO owns the response.
Give agent management its own place on the manager development roadmap. Engagement training will not cover it, and neither will training focused on
supporting people’s use of AI tools. Both still have utility but fail to
deliver on the basics of effective agent management.
Create the capacity. A manager who runs agents well needs hours the current job does not leave. In my build, the early investment ran to 20 hours in some weeks. Plan for that in workloads, or the loop never runs.
Hold a standard for who takes the role. Diagnosing, adjusting, and checking works for someone who wants to do it and keeps doing it after the novelty fades. People who accept a manager title for status or a raise will not. Gallup estimates companies miss on high managerial talent in 82% of their hiring decisions. I know many CHROs do not choose many of those managers, and that a team in urgent need tends to get whoever is closest. Build the bench and the selection criteria before the urgency arrives and ask your CEO to hold the line.
Start with yourselves. CEOs and CHROs are managers too. Build one small agent. Grade it on a schedule. Write down every change you make and what the next report says about it.
Agents compress the management feedback from years to days. Your first D isprobably yours, like it was mine.
Sources
Photo Credit: Pexels, Robert So