Why managed agents are the next big thing in agent building

Jessica So, managed agents are officially the new hot thing now. Harrison just did the whole “this is the next phase of agent building” post around Managed Deep Agents.

Cathy Yeah. And for once it’s not just vibes, there’s an actual argument there about harness plus infrastructure bundling.

Jessica Before we get nerdy… how’s your week going? You sound less doom-scroll and more “I finally got sleep process scheduling under control.”

Cathy I mean, in as much as I ever do. I’ve basically spent three days mentally refactoring every bad agent harness we’ve seen into this business-logic, harness, infra split.

Jessica That is such a you hobby. I just read it and went, okay cool, fewer teams setting themselves on fire trying to run AutoGPT clones in production.

Cathy Right.

Jessica So the core claim here is pretty specific. He’s saying: we had early frameworks like LangChain and LangGraph, then harnesses like Deep Agents and Claude Code, and now we’re in this “managed agents” phase where the harness runs on managed infra and you steer it through standards like AGENTS dot md, MCP, and skills.

Cathy Yeah, and he actually walks the timeline. Late twenty twenty-two to early twenty twenty-three is frameworks and apps. Then early to mid twenty twenty-five, models get good enough that the default pattern becomes “LLM in a loop calling tools.” Manus, Deep Research, Claude Code all look like that loop.

Jessica And then the harness moment—Claude Code, Pi, Deep Agents—where people stop hand-wiring loops and start reusing patterns. That’s literally our whole harness-as-product rant from like four episodes.

Cathy Exactly. Then the two learnings he calls out are what make the managed story feel real: infra primitives and control standards. On the infra side, you get durable execution for the loop, sandboxes for untrusted code, this “separate brain and hands” design that keeps the L L M away from raw system access.

Jessica Mm-hm.

Cathy On the control side, you get AGENTS dot md as the canonical instruction file, MCP for plugging into external systems, skills for progressive disclosure of context. Once those are semi-stable, you can say, okay, we’ll run the harness for you, you just send us files and configs.

Jessica Which is the managed agent pitch: you always own your business logic—your tools, your context, your instructions—but instead of picking a harness and then solving runtime, streaming UX, sandboxes, context hub, eval, memory, auth one by one, you pick a bundle where those are pre-wired.

Cathy He even lists them like a bug tracker. Runtime reliability and resuming mid-run, streaming events back to a UI and “bringing the agent to where users already are,” where code executes, where context lives and whether subject experts can edit it, how you eval changes, what the agent remembers, and who it’s allowed to act as when it hits external systems.

Jessica As a product person, that list reads like “all the things your P M quietly panics about once the demo works.” It’s exactly the boring stuff I get excited about.

Cathy You do love boring stuff when it ships.

Jessica I do, and this is also them owning their own evolution. Fleet was the first managed swing—no-code, pure UI, aimed at non-developers, with agents literally represented as files in a virtual file system. Toggle to see AGENTS dot md, skills, whatever, as files the harness loads.

Cathy And the punchline was: their actual users are developers, so Fleet generated a wall of requests for custom middleware, custom tools-as-code, and API-based agent creation. Which is basically devs saying, “please keep the infra managed but stop locking me in the Lego Duplo editor.”

Jessica Then you get Claude Managed Agents: still a UI, but really A P I first, more obviously for developers. Similar standards, less file-system cosplay. Plus that “dreaming” concept, which is its own thing.

Cathy And Vercel Eve, leaning super hard into agents-as-files and bundling deploy infra. So by the time he says, okay, today we’re launching Managed Deep Agents, it doesn’t feel out of nowhere. It’s another entrant in that same managed harness bucket.

Jessica Let’s talk that bundle specifically, because this is where I think it’s most useful. It’s Deep Agents the harness, agents-as-files for representation, plus knobs for custom middleware and tools-as-code. Then infra-wise, it’s LangSmith Deployments for runtime, streaming from the LangSmith Agent Server and Channels for UX, LangSmith Sandboxes, Context Hub for context management, Harbor for evals, opinionated memory on top of Deep Agents, and built-in auth.

Cathy So basically, their whole “boring controls win” story wrapped around the harness. From a systems angle, that checks out. Durable execution plus sandbox plus eval plus auth is the minimum viable skeleton for long-horizon agents that can touch real systems.

Jessica Do you buy his “this unlocks a new wave of production agents” line, or is that just launch-day adrenaline?

Cathy I think it’s… directionally right, with caveats. If your product is one-shot question answering or a single tool call, this is overkill. You don’t need a deep harness or managed runtime; you need a good prompt and maybe a router. But if you already committed to agents that run in loops, call tools, and maybe run code, then yes, the infra is what’s killing you, not the model.

Jessica So the real “who should care” is teams that already know they want that loop: multi-step support workflows, deep research, coding agents hitting repos and sandboxes, that kind of thing. Not “I added an L L M to my form.”

Cathy Yeah. For those teams, the argument holds up technically. The business-logic, harness, infra split is clean. Bundling harness plus infra into a managed service is a real efficiency win, especially if you were going to build a bunch of this yourself anyway.

Jessica My only hesitation is the standards bit. AGENTS dot md, MCP, skills are still kind of an ecosystem handshake more than an I E E E spec. If you go all in on Managed Deep Agents, how portable is that agent to Claude Managed Agents or Eve?

Cathy That’s my main open question too. He’s honest that we’re early and more infra and standards will show up, which I appreciate. But the portability story is not solved just because everything is “files.” It’s still their harness’s interpretation of those files.

Jessica Still, as someone who just wants fewer teams inventing their own eval harness and half-broken sandboxes, I’m pretty into this. It slots almost perfectly into our “harness-as-product and boring tooling is load-bearing” thing.

Cathy Yeah, this is one of the rare posts where your product optimism and my harness skepticism end up nodding in the same direction. I think the central claim is basically right: managed agents are the next stage for people who are already playing the agent game.

Jessica Okay, Cathy, I’m going to go rewrite half our imaginary backlog to assume “managed agents exist now” and you can tell me later which parts are fantasy.