Anthropic launched Claude Managed Agents on April 8. The announcement reads like a developer update.
It isn't.
It's a new build-vs-buy question landing on every PM's desk.
What Managed Agents actually does
Until now, running agents in production meant building infrastructure first. State management, permissioning, tool definitions, security review — the engineering work that had to happen before the business work. Teams routinely spent three to six months just getting to launch.
Managed Agents says: skip that part. File reading, command execution, web browsing, code execution — all running in a managed environment Anthropic controls. You get from prototype to production in days.
Early numbers are real. Rakuten deployed specialist agents across sales, marketing, finance, and HR — all live within a week. Asana's CTO said they shipped features "dramatically faster" than before.
Those aren't demo metrics. That's production.
But speed isn't the whole story.
MIT research says 95% of AI pilots deliver no measurable business impact. Gartner predicts over 40% of agentic AI projects will be scrapped by 2027.
So speed matters. A lot. Projects that never leave the pilot phase are, by definition, failures. If managed infrastructure is what gets you to production, that's a legitimate business argument.
Here's the thing though: speed compounds lock-in.
Vendor lock-in isn't abstract anymore
When you run agents on a managed platform, the workflow doesn't just touch the platform — it lives there. State management, permissioning, monitoring, security. All of it embedded in Anthropic's system.
You only notice this when you try to leave.
What happens if the platform triples its price in six months? How much does migration cost when your agent's operational logic is woven into someone else's infrastructure?
I'm not raising this to be alarmist. I'm raising it because PMs don't usually ask this question until they have to.
In healthcare SaaS, we see this pattern constantly. Connect to an EHR integration or an e-prescription system and you're not just moving data — you're handing over workflow logic. Leaving later isn't a technical swap, it's an operational transformation.
"Managed" is a strategic word, not a technical one
When most PMs hear "managed infrastructure," they route it to engineering. But "who manages this?" is a product decision.
The two positions are clear:
Build your own — full control, portable, but six months minimum and serious engineering capacity required. High risk of never shipping.
Use a managed platform — production in days, but lower control and weaker portability.
For most teams — especially those without dedicated platform engineers — option two is the right call. The logic holds.
But it should be a conscious call. Not a default.
A few questions worth asking before committing: What's your exit strategy if pricing changes? Who owns measurement — you or the platform? What happens to your workflow data if you need to switch providers a year from now?
Build-vs-buy used to live in the software category — do we buy a CRM or build one? It's migrated to AI infrastructure now.
This is shaping up to be one of the most consequential PM decisions of the next two years. Not because the technology is complicated. Because the organizational commitment is deeper than it appears.
In your company, who's making this call — PM or engineering?