--- title: "The Scaffolding That Used to Be Built In-House Is Now for Rent" date: 2026-11-05 category: Foundation Models site: NeuroAI canonical: https://neuroai.site/a/na-hosted-agent-07-1105b language: en --- # The Scaffolding That Used to Be Built In-House Is Now for Rent > On November 5, the same day it released the model weights, Moonshot AI also launched a service called Kimi Hosted Agents, known in Chinese as hosted agents. The name is a mouthful, but the idea is On November 5, the same day it released the model weights, Moonshot AI also launched a service called Kimi Hosted Agents, known in Chinese as hosted agents. The name is a mouthful, but the idea is not: the back-office frame that every company used to build and mind for itself has now been tidied up and put on offer to outsiders. It helps to picture an agent as a helper who can keep working for a long stretch. Unlike a single question and answer that ends as soon as it is spoken, it has to look up material, write code, and call tools, and pick up again if interrupted. One job may run for hours or longer, so there has to be somewhere for it to sit steadily. That somewhere is exactly what this service rents out. The trade term is Harness, and this article simply calls it scaffolding: it is not the model itself, but the whole underlying environment that holds the model up while a long job gets finished. The model does the thinking; the scaffolding keeps it from falling off halfway. ## The Scaffolding Hidden in the Back Office What does the scaffolding actually manage? According to The Paper, under the traditional model, a company that wants agents to run steadily usually has to build its own code execution environment and also settle a string of chores: context management, tool calling, permission control, and recovery from interrupted tasks. None of them stands out, and none can be spared. Taken one by one, it becomes clear. The code execution environment is the place where the helper can actually do hands-on work; context management is what lets it remember how far it got; tool calling opens the drawers it is allowed to touch; permission control marks which things may be touched and which may not; and recovery means that when a job breaks off halfway, it can pick up again. These jobs share one trait: they have little to do with the business itself, yet nobody can skip them. To have an agent do serious work for a company, someone first has to lay this back office brick by brick and then keep it up over time. For many teams, the hard part is not coming up with the idea but pouring a solid foundation first. Nor should scaffolding and model be mixed up as one thing. The model is the brain that thinks and answers; the scaffolding is the desk, the notebook, and the duty roster holding that brain up. Switch to a cleverer model and the desk still has to be fitted out again. What this service rents out is the latter, not the brain thrown in along with it. ## Building It Yourself Means Several Hurdles Building this frame costs more than the initial burst of effort. Tools get updated, permissions have to be adjusted along with them, and long-running jobs throw up all sorts of unexpected snags; what connects today may not connect a while later, so someone has to keep watch and keep fixing, without letting go. What is more awkward is that it is not like buying a piece of equipment that is done once installed. How context is stored, where an interrupted job resumes, and how several helpers working at once avoid colliding are all fine details to be ground down through use. According to The Paper, once a company builds such reliable infrastructure, it still has to keep putting in maintenance costs afterward. And so a familiar situation arises: an idea is not far from thought to execution, yet simply getting an environment ready for long-running work consumes most of the energy up front. What holds up progress is often not that the model is not clever enough, but that the frame holding it has not yet been built steady. ## Moonshot Says It Could Take Months Just how much trouble this frame is has an answer from the company itself. Moonshot AI says that if a company builds a reliable Harness infrastructure on its own, it may take several months, with continued maintenance investment afterward. Note that this is an estimate given by Moonshot AI, not a figure worked out by a third party. The estimate of several months is worth pausing over. It refers to a complete environment built from scratch and able to run reliably over the long term, not to writing a small script for a trial. Read those words together with the maintenance cost that follows, and it becomes clear exactly which stretch this service is trying to spare people. Precisely because it is the company's own account, it is best read with a sense of proportion. It is enough to explain where the service came from; as for how long any given company really needs, team size and business complexity differ, and no one should take this figure and use it to schedule or budget someone else's project. ## Two Kinds of Use Cases Are Listed First According to the official introduction, the service is aimed mainly at two kinds of use cases. One is integration into a company's internal systems, taking on back-office work such as business monitoring and operations response. The other is embedding it into products a company puts in front of users, serving as an investment-research assistant or a customer-service agent that can answer directly. The two look different but rest on the same foundation: both need the helper online for long stretches, both need a task to be resumable midway, and neither requires the company to mind a machine room itself. One faces inward, watching systems for staff; the other faces outward, answering questions for customers. The scaffolding is still the same scaffolding. No examples are being filled in here on anyone's behalf. The material gives no specific customers, says nothing about how many people use it, and gives no price, concurrency figures, or service commitments; what is not written down does not get added. What can be settled is only the two kinds of use listed officially, and the running environment both of them need. ## Companies Just Assign the Work Once it becomes a service, the division of labor changes. According to The Paper, a company only needs to define tasks, configure agents, and connect its business tools, while Kimi handles maintenance and upgrades of the underlying running environment, so as to shorten the cycle from prototype development to actual deployment. The official account also mentions that the environment is hosted around the clock, supports long and complex tasks, and supports collaboration among multiple agents. In plain words: you used to build the kitchen and mind the stove yourself; now someone watches the stove, and the company simply hands in the recipe and ingredients and waits for the dish. With that change, what a company has to worry about shifts from how to build the frame back to how to assign the work. Which tasks go to the helper, at what point a person should take a look, and who to turn to when something goes wrong — those judgments still rest with the company itself. What the service takes away is only the layer of repetitive labor underneath. ## What Gets Saved Is Time, Not People One thing should be said up front: what this service sells is saved time, not decisions made on anyone's behalf, still less the removal of people from the process. It bundles up the repetitive work of building and minding the frame, while judgment, review, and responsibility still stay with the team using it. For ordinary readers, none of this is unfamiliar. Just as not everyone in a household needs to know how to fix a pipe in order to turn on a tap, handing specialized, fiddly underlying work to a specialized party and freeing your own hands for what matters more is simply the most ordinary shape of division of labor. So what is worth remembering about this service is not another English name but one sentence: even the scaffolding that runs long tasks has begun to be treated as something that can be rented. It used to be hidden in every company's own back office; now it has been moved onto the counter. **Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.** --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-hosted-agent-07-1105b Free to quote with attribution and a link to the original.