← Analysis

Analysis · 2026-11-13 · By bombong

What changes when agent execution environments become cloud products

This week, Google Cloud, Moonshot AI, AWS and Mistral AI announced harnesses, isolated sessions and runtimes for running agents as managed services, or published ways to run them. This article lays out what the execution layer contains and how companies differ on where it runs. It then shows how lock-in and cost structure are extending beyond the model API.

A large warehouse lined with glass rooms, each with a small robot working inside; some rooms are dark and closed, and at the edge, people stand at a fork in the road.
AI-generated image · A large warehouse lined with glass rooms, each with a small robot working inside; some rooms are dark and closed, and at the edge, people stand at a fork in the road.
Key points

What came out all at once this week?

What this week's agent news had in common was not the models themselves but the execution environments that run them. At the center is the harness (Agent Harness). A harness is the execution framework that attaches tools, memory and a sandbox to a model so it can actually do work. Several announcements selling this harness as a managed service came out in the same week.

According to AI Times Korea (인공지능신문), Google Cloud made Agent Substrate 1.0, an execution platform for autonomous agents, generally available on the 5th, US time. It also open-sourced three pieces of operational infrastructure source. The core feature saves the state of idle agents and restores it when they are called again source. Google said execution density is up to 10x higher than conventional container approaches and saved state is restored within 0.5 seconds. However, the comparison conditions were not disclosed source. According to Pandaily, Moonshot AI launched Kimi Hosted Agents, a managed harness that runs 24 hours a day. It comes with a sandbox, audit logs, Inference Hook and eight financial templates source. Mistral AI announced that it added managed agents and apps, plus agent observability and evaluation features, to Studio source.

Demand and adjacent markets moved too. According to Pandaily, DeepSeek Harness has 1.54 million daily users and processes about 92 trillion tokens a day source. TechCrunch reported that Dropbox is acquiring Tigris Data for $350 million to build DBX Agent Cloud, storage for agent developers source. With cloud providers, model companies and storage companies all moving on the same layer, the point of competition can be seen as widening, from "which model to use" to "where and how to run it." This means that when evaluating agent adoption, model and runtime now need to be chosen separately.

Numbers around the execution layer

Up to 10xExecution density claimed by Google (vs. containers)
0.5 secondsState restore time claimed by Google
1.54 millionDeepSeek Harness daily users
$350 millionDropbox's acquisition price for Tigris Data
Google's figures are company claims, and the comparison conditions were not disclosed.

What does the execution layer contain?

The execution layer sounds abstract, but its components are concrete. A guide published by AWS is a good example. AWS explained how to run Claude Code as a remote agent on Amazon Bedrock AgentCore. A separate isolated session is created inside the customer's AWS account for each coding task. All model calls go through Amazon Bedrock, and results come back as GitHub Pull Requests that a person reviews and merges source. AWS also gave reasons for choosing this setup. It is billed to the existing AWS account without per-seat licenses. Audit logs are recorded in AWS CloudTrail. ZDR (zero data retention), which does not store requests and responses, can also be configured source.

The same trend appears on the hardware side. NVIDIA pointed out that agents switch between GPU inference and CPU execution during a task, explaining that search, DB queries and code execution are steps handled on the CPU source. NVIDIA said partner Daytona launched 1,000 agent sandboxes on the Vera CPU in 10.8 seconds. On x86 it took 88.4 seconds, according to NVIDIA source. Storage has also begun to be treated as part of the execution layer source.

Taken together, agent operating costs do not end with model call fees. Isolated sessions, storage of idle state, audit logs, CPU execution and storage can all be seen as adding to the cost. Google's emphasis on state saving and execution density also reads as targeting this idle cost source.

Remote coding agent flowDeveloper → Isolated session(Delegate task); Isolated session → Amazon Bedrock(Model calls); Isolated session → GitHub PR(Return result); GitHub PR → Human review & merge(Review)DeveloperAssigns taskIsolated sessionInside customer AWSaccountAmazon BedrockAll model callsGitHub PRTask resultHuman review &mergeDelegate taskModel callsReturn resultReview
Remote coding agent flow The setup in the AWS guide: each task runs in an isolated session inside the customer account, and results come back as a PR for human review.

Why do companies differ on where to run it?

Even for the same execution layer, who operates it and where it sits differs by company. At one end is building it yourself. According to a video summary shared on GeekNews, GitHub ported roughly 400,000–450,000 lines of its TypeScript Copilot runtime to Rust. The goal was a common agent harness that could be embedded in multiple products without Node.js and V8 source. The switch was made gradually over about 14.5 weeks without halting deployment, with about 130 releases shipped during that time source.

Mistral AI said Studio can be deployed in the cloud, in an enterprise VPC, on-premises, or in a mix of these. Mistral's post also included a comment from a BNP Paribas representative, who said the bank runs Studio on its own infrastructure because it handles sensitive processes source. AWS offered a middle ground: execution is delegated but kept within the customer account boundary source. Moonshot AI's service sits at the end where teams do not need to host the runtime themselves source. Google released both a managed platform and open-source infrastructure, but the names and licenses of the three released components have not yet been confirmed source.

This difference reads less as a matter of technical taste than of differing control requirements. Organizations for which data sovereignty and auditing matter may choose the inside of the boundary, while those looking to save on operations staff may lean toward fully managed. Deciding first where your organization should sit on this spectrum quickly narrows the candidates.

Who runs the execution layer?

← Build and run yourselfLeave to provider →
12345
  1. GitHub shared runtime
  2. Mistral Studio self-hosted deployment
  3. AWS isolated sessions in customer account
  4. Google managed + open source
  5. Kimi Hosted Agents
Further left means you build and run it yourself; further right means you leave it to the provider. Positions are relative, based on the article's interpretation.

How do lock-in and costs change?

Model APIs are relatively easy to swap out. The execution layer, however, accumulates a lot: saved agent state source, audit logs sourcesource, work templates source and plugins. According to Pandaily, DeepSeek Harness has about 35,000 plugins built by outside developers, and an official marketplace and creator program are in preparation source. As these assets grow, lock-in may shift from the model to the harness and its ecosystem.

The cost structure changes too. AWS itself listed the trade-offs of remote coding agents: less intervention during execution, the need for environment setup, and costs that are harder to predict source. It also distinguished which tasks fit and which do not. Good fits are scoped features, bug fixes and test work in trusted repositories that have their own setup and test commands. Untrusted code is not a good fit, it said source.

In short, comparing model prices alone is not enough when choosing a managed runtime. Session, idle and storage costs, along with whether accumulated assets can be moved elsewhere, can also be considered decision criteria. How much an open-source release source or multi-environment deployment source will improve portability is safer to judge after checking the licenses and the actual components.

When to hand off, when to hold on

Hand off to managed harnessReduce operational burden

  • No need to host the runtime yourself
  • Use templates and plugins right away
  • State and logs accumulate in the provider's environment

Keep within your boundaryRetain control

  • Self-hosted, VPC or on-premises deployment
  • Direct control over data and model access
  • Setup, operational burden and cost predictability are challenges

공통 Either way, the execution layer outside the model shapes cost and lock-in

A comparison of handing work to a managed harness versus keeping the execution layer within your own boundary.
What practitioners can try
  1. Split the cost of the agent workloads you currently run into 'model calls' and 'isolated sessions, idle state and storage', and compare the two items separately when you get quotes for a managed runtime.
  2. Before signing up for a managed harness, check whether agent state, audit logs, and templates and plugins can be exported to another environment, and what licenses the open-source components use.
  3. Start remote coding agents on scoped feature additions, bug fixes and test work in trusted repositories, and let results be merged only after PR review.

Related terms

Harness Sandbox AI agent Coding agent Always-on agent Human-in-the-loop LLM observability Sovereign AI

Sources

This piece was drafted with AI, checked against the cited public sources for links and figures before publication, and reviewed again by the operator afterwards. It includes the author's interpretation. How we write and check · Corrections