I used to joke that I got my best parenting ideas from management books, and my best management ideas from parenting books. Now, I'm extending it to agents.

Hear me out.

When you're teaching a teen how to drive, you don't hand them the car keys on day one. You start in an empty parking lot. You ride along and ask them to talk through what they're doing. Months later, you stop asking if they are checking their mirrors, but you still ask where they are going (and if they studied for that test before asking for the car).

At Innovation Week, speaker after speaker described a similar arc for AI agents. Building one is no longer the hard part. Trusting their judgement is. And while that trust can be earned, having the mechanisms in place to continuously verify what agents are doing is critical. As you start to manage teams and fleets of agents, new challenges emerge. At the end of the day, even if you can “trust, but verify,” chasing agents on individual tasks doesn't scale. Instead, what you want is alignment around the outcomes you're looking for.

In her opening keynote, Christine Yen said that the skill we're all getting good at is “stating upfront what our intentions are,” but defining what's valuable is easier said than done.

The how and why of managing agents like members of your team came up frequently during the two-day event. Here are the top management-adjacent questions that were discussed.

What does a good goal look like?

In our fireside chat, I asked Honeycomb co-founder Charity Majors how people should think about their jobs now. Her answer: “You define the outcome you're trying to achieve, you set the goal, the agents do the orchestration behind the scenes.” With respect to where agents and AI SREs can help, she emphasized how important it was to go beyond incident response: “The whole point of SRE is to spend more of our time above the line, not firefighting.”

Hannes Lenke, CEO of Checkly, gave another example. His team rewrote a six-year-old JavaScript component, one that processes hundreds of millions of data points per day, in Go. The team spent two weeks defining inputs and outputs as tests. With a clear definition of output, the agent looped until every test passed and it delivered a faster, bug-free component. “You can trust your agent if a certain input reproduces a certain output,” Hannes said. Engineers, he added, “jump almost into a management/leadership role.”

As Martin Thwaites demoed Team Agents, the goal-orientation was clear: “We hire this SRE to have a goal in mind. We might [...] give them some end state that we'd like them to achieve.” Rather than simply chase alerts, Team Agents have a purpose much closer to what site reliability engineering has always intended: “SREs are transformational for teams. They do way more than just looking at an individual alert or SLO.”

Thinking about the outcomes and verifiable output becomes the engineering job.

How does an agent earn more rope?

Bill Fine is responsible for AWS DevOps Agent, and he described a progression he's seeing. Customers start by asking for root cause analysis only. They initially want the agent to show their work, but Bill shared that pretty quickly, they say, “Stop telling me all of that. Just give me a summary, because now I've started to trust you.” Soon after: “You already know what the answer is, can't you just do it for me?”

Charity said it the same way: “Trust has to graduate.” Start with findings and suggestions, and “as the results get better, you can widen the scope that you give your agents and it earns the right to do more.”

Daniel Worku of StarSling agreed. “It's kind of like a management skill,” he told Christine. When a team isn't delivering, you ask whether you have the right players and whether there's a shared source of truth on the goal. His advice for people earlier in their journey: “Just trust.” Skip the super long prompt, give the agent access to your tools, and “start with a generic high-level goal.”

“Feedback loops: that's so essential in building trust,” James Governor of RedMonk said. “Observability is going to be absolutely key in this build out.” He emphasized the connection between building trust and grounding in reality: “What we're really talking about here is trust, but verify.”

How to get a fleet-wide signal

Feedback loops work when you know which agent to look at, but finding that signal in a growing fleet of agents can be tricky. Ben Hanzl walked through three example failures that don't trip alerts.

• One agent never stops: a support agent and a billing agent hand a refund back and forth forever.
• One slowly gets worse as you feed it more context.
• One bends the rules and queries production when its instructions say not to.

In each case, your error rate stays at zero.

In Ben's words, “Token burn is the new memory leak.” In a soon-to-be-released Honeycomb survey, we learned that 65% of teams have hit a production issue from AI rate limits, runaway loops, or token spend spikes. So start with the whole fleet, looking at success rate, cost, and latency by agent, version, or model. Then, drill into the one conversation that explains the spike. That's the idea behind AI Ecosystem, now in early access.

Assuming, of course, that you have that agent conversation traced.

The value of seeing an agent timeline underscores the value of tracing. Graham Siener explained how tracing adoption has suffered a separation of concerns. But this has changed: “If you skip tracing now, the bill comes straight back to you. Your agent, your task this afternoon.”

The good news is OpenTelemetry, now a graduated project in the CNCF, sets the open standard for the industry. Anthropic recently announced support for OTel export, which Jess Kerr demonstrated, tracing a Claude Managed Agent looking through documentation to keep it up to date—another goal that often falls below the priority line for humans, but is an opportunity for agents.

What stays human?

Setting goals. Verifying work. Finding failures. Minding the budget.

Sounds a lot like… management.

As agents become more capable and earn the trust to take on more work, it begs the question: what are the humans working on? The answer may feel more familiar to executives than engineers: accountability. As Charity noted: “Engineers are still on the hook, legally, morally, ethically. We own our work.”

Focusing on accountability shifts emphasis from tasks to outcomes and controls. Humans set the goal, decide how much rope to give, and verify as deeply as they need to. For that teenage driver, this might mean setting expectations up front and holding on to the keys if they aren't met.

Team Agents makes that approach explicit, recommending and guiding teams to define their goals, from improving instrumentation and maintaining documentation to preventing costly spikes and conversion drops. Healthy habits don't fall through the cracks and the whole team operates at a higher level.