Anthropic shipped Claude Sonnet 5 on June 30, 2026 and called it the most agentic Sonnet model yet. It plans, picks up a browser, runs a terminal, and works through a multi-step task without someone checking in every five minutes (Anthropic's own announcement). I read the launch post the same week I was rewriting a set of one-on-one notes for a manager who still approves every ticket before an engineer touches it.
Two documents, same week, opposite philosophies. One trusts a system to plan its own work inside a boundary. The other doesn't trust a senior engineer to pick which ticket to start.
I've led engineering teams for twenty-plus years, and I build Step Up To BAT, a tool built around one idea: managers are the biggest lever in an organization, for better or worse. So when I sat down with Anthropic's engineering notes on building effective AI agents, I kept swapping "agent" for "employee" in my head. The fit was too tight to ignore.
What Anthropic Tells Its Own Engineers
Anthropic's guidance on agent design isn't a hype piece. It reads like a discipline. Three lines stood out to me.
First: start simple. Anthropic tells developers to add complexity only when it demonstrably improves outcomes, not because a bigger system sounds impressive. Most tasks don't need an agent. They need one clean, well-scoped function call.
Second: build a real interface, not a permission form. Anthropic treats the agent-computer interface with the rigor most companies reserve for the customer-facing product. Tools get documented, tested, and refined the way you'd refine a UI a paying customer touches every day.
Third: guardrails replace supervision. Anthropic doesn't recommend watching every step an agent takes. It recommends testing in sandboxed environments and setting boundaries up front, then letting the system operate inside them. Autonomy is the point. Surveillance isn't a substitute for trust. The sandbox is.
Put those three together and you get a working definition of a well-managed agent: a narrow, well-documented job, real tools to do it, and a boundary instead of a babysitter.
Now read the list above again and ask how many of your direct reports get all three.
The Manager Who Never Learned to Let Go
I've been the manager who checked every ticket. Early in my career I thought oversight was the job. I reviewed every pull request line by line before it merged, not because the code was bad, but because I hadn't worked out how to trust a system I didn't personally watch every second of.
It didn't scale. It didn't work either. My best engineers started routing around me, quietly, the way any competent system routes around a bottleneck. The ones who stayed stopped bringing me their hardest problems, because bringing me a hard problem meant a week of me re-deciding it for them.
Anthropic's own agents get more latitude than I gave a senior engineer with a decade of experience. Sit with this for a second.
This Is Psychological Safety Wearing a New Badge
None of this is new. It's old research in new language. Amy Edmondson spent decades showing teams which feel safe raising problems outperform teams which don't, because oversurveillance and constant second-guessing raise the social cost of speaking up and shrink the space where anyone learns anything (Psych Safety's breakdown of the research). Self-determination theory says the same thing from a different angle: strip someone of autonomy and their motivation goes with it.
Anthropic didn't invent the idea a system does better work when it's trusted with a clear goal and left to find its own path inside a boundary. They rediscovered it, wrote it down for AI agents, and now every engineering team building on Sonnet 5 is quietly relearning a lesson management researchers have written about since long before the first LLM shipped.
Why This Is Landing Right Now
Gartner predicts 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025 (Gartner's own prediction). This is an eightfold jump in a single year. Every engineering leader reading this is going to spend the second half of 2026 designing systems meant to operate with real autonomy inside real boundaries.
Here's the uncomfortable part: most of those same leaders haven't designed a boundary like it for a human on their team. We're about to get sharp, and fast, at writing clear goals and real guardrails for software. It would be a waste to not apply the same discipline to the people who built the software in the first place.
What Good Agent Design and Good Management Share
Break the parallel into three practical habits, the same three Anthropic bakes into its agent guidance:
Set the goal, not the steps. Tell an agent what "done" looks like and hand it the tools. Tell a person the same thing. Micromanaging the path defeats the point of hiring someone smart enough to find one.
Invest in the interface, not the paperwork. Anthropic spends real engineering effort making tools legible to the agent using them. Most companies spend the effort on approval chains instead. Give your team clean systems and real access, not a form to fill out before they're allowed to act.
Build the guardrail before the task starts, not the check-in after. A sandbox with clear boundaries beats a manager reviewing every step after the fact. Decide up front what "off the rails" looks like, then get out of the way until it happens.
Autonomy Isn't the Same as Absence
Anthropic doesn't hand every agent a blank check. Their own guidance is blunt about it: human oversight stays essential in high-stakes domains, coding and customer support among them. Trust the system inside its boundary, but keep a human in the loop where a mistake costs something real.
This is the part of the parallel most leaders skip. Removing yourself from every decision isn't the goal. Removing yourself from decisions which don't need you is. A senior engineer picking which ticket to start on this sprint doesn't need your sign-off. A senior engineer about to touch the payment system on a Friday afternoon might still be worth a second set of eyes, the same way Anthropic still reviews its highest-stakes agent actions before they run. Autonomy scoped to the risk in front of you, not autonomy across the board, and not surveillance across the board either.
Where I've Put This to Work
At Step Up To BAT, the premise is simple: a manager is not able to fix a blind spot they don't see. The feedback loop does the work oversight used to do badly. It shows a leader where the boundary works and where it operates as surveillance dressed up as support. I wrote more about the mechanics on Step It Up HR, if you want the deeper version.
The pattern holds whether the "agent" runs on a GPU or draws a paycheck. Give it a real goal. Give it real tools. Set the boundary once, clearly, and stop reviewing every step inside it.
The Question Worth Sitting With
Anthropic didn't build Sonnet 5 to be watched every step of the way. They built it to be trusted inside a boundary they took real time to design.
Look at your own team the same way. Did you spend more time this month writing clear goals and real access, or writing approval steps and check-ins? Anthropic's agents get more trust than most managers give their best people. Fix it before you ask why your best people started acting like they don't trust you either.