Claude Managed Agents, as three founders actually use them

Anthropic is the AI company that makes the Claude models, and Managed Agents is the agent-operations service it launched in April 2026. The video is less a product advertisement than a session that asks what each of the three companies decided to build themselves, what they decided to rent, and where they drew that line.

To put it simply, Managed Agents is a service in which Anthropic runs the servers, memory stores, and isolated workspaces needed to operate an AI agent on your behalf. It is like opening a restaurant by renting the kitchen and the storeroom so you can concentrate on the cooking.

Wispr Flow uses the service for a meeting-prep briefing feature. Twenty-four hours before a meeting, an agent prepares a summary of who you are about to meet, why, and what you need to get out of it. Founder Sahaj Garg says that if the brief is wrong, it is worse than having none at all. If you are meeting someone named Tom and it shows you information about the wrong Tom, the user ends up confused in the meeting itself.

That is why the feature the company values most is Outcomes. You give the agent a grading rubric, and a completely separate agent from the one that did the work grades the result against it. The rubric mixes objective items, such as whether it pulled the right person's LinkedIn, with readability items, such as whether the material is laid out to be skimmed at a glance rather than delivered as five pages. If the grade fails, the company chooses to show the user nothing.

Actively assigns one agent to each customer account. If a sales rep covers 100 accounts, 100 agents run in parallel, each building up its own memory. Questions that cut across accounts, such as which five accounts to go after today, could not be answered in that structure. Founder Mihir Garimella says he built a feature called Watchtower to fill that gap with Managed Agents in two weeks, and it has since become one of the most used features.

What this company rents is the memory store. Core memory for each account is managed in-house because of retrieval speed and corruption prevention, but the rest, such as how an organization does forecasting or what a user prioritizes, is left to the managed memory the service provides. Garimella's explanation is that there is no reason to build anything yourself that does not need to behave differently from everyone else's.

Pendo's use case is a little different. The company sells product analytics tools that record what users actually do inside a customer's app, and its new agent also takes in the customer's source code and matches what the code intends to do against what actually happens. Every time a pull request goes up, the agent reads the code and does a light usability review, and at night it checks whether the previous day's conversion rate dropped, tracing back as far as changes made four weeks earlier to propose fixes.

Because it handles other people's source code, what Pendo rents is the sandbox, that is, the isolated workspace. CEO Todd Olson says that the moment you touch source code, secrets can leak, and you have to think carefully about which tools you expose and which data you send, none of which is the company's core value. They originally had a system they had built themselves, but as problems piled up they compared it with Managed Agents, and he says the infrastructure was running within days and the whole product had moved over within weeks.

All three apply the same test for whether to build it themselves: does the infrastructure determine the quality of the feature you are building right now. Garg says that for the briefing feature the framework that runs the agent is not a core competency, but if he were building a fully voice-controlled assistant, that framework would be the product, and he would most likely build it himself. Olson adds that in the early stage no company is an infrastructure company, and getting features in front of customers fast comes first.

They do not hide the complaints of the renting side either. Garimella says that when he fans work out across 500 accounts he cannot use the top-end model and has to switch to a cheaper one, but the service's structure is close to all or nothing, leaving few levers to cut costs. Olson says that if work that could be run ahead of time overnight could be handled in batch mode, he could save between half and three quarters of the cost, but that is hard today. The Anthropic host responds that a batch mode would help.

The conversation about evaluation is candid too. Garg says they built by feel at first, but from the moment the product started to click they needed an evaluation set, and Garimella says the approach that worked best was turning questions real customers had asked into evaluation problems while talking with those customers. There is also an admission that they have not yet found a way to evaluate offline a system whose memory keeps changing, or an agent that depends on outside services like Slack.

They also point to the trap that comes with each new model. Garg says the Claude 5 series produces more sentences that read as AI-written than the previous generation and uses em dashes more often, and that the most valuable thing Anthropic could do is warn them in advance about the new failure modes a new model brings in. It is a scene in which the head of a company building its own product points out the habits of a model company's model, in front of that company.

Where this video speaks to Korean developers is in the arithmetic rather than the conclusion. All three companies knew that building it themselves would be cheaper and freer, yet chose to ship features in a day or two weeks, and they put numbers on what they gave up in cost and control in return. A team debating whether to write its own agent infrastructure can work out its own answer just by watching the order in which these three answer the same question.

In the end, what Anthropic wants to show with this video is that Managed Agents is fast, and what the three founders add is the price of that speed. The shortest answer to come out of the day: if the cost of experimenting is the time it takes to learn, renting is right, and if a feature only works when you cut costs by 90%, you have to build it yourself.