#architecture#team

Why Laptop Agents Fail: The Case for a Cloud Harness

Learn why coding agents on laptops fail to scale and what cloud infrastructure requires: state, sandboxes, access controls, and limits for IT leaders.

  • What Latent Space analyzed
  • What is a harness
  • Meme about an open laptop
  • What must be solved during migration

What Latent Space analyzed

Latent Space analyzed why coding agents are so difficult to move from a developer's laptop to the cloud (original source). The main point is this: while the agent harness lives on one machine, the company owns an employee's personal tool. It starts owning a service when the harness runs in the cloud and the operations team manages access, logs, and the budget.

What is a harness

A coding agent has a simple architecture: in a loop, the language model calls tools and carries along the necessary context. The entire loop, together with its tools, permissions, and memory, is called a harness—the layer surrounding the model. It is easiest to build locally, which is why most agents started in a terminal and later moved to desktop applications. According to the authors, Claude Code also began as a CLI utility.

Meme about an open laptop

  1. Local execution has a clear limit.

  2. The agent works on a task for an hour or two, the developer closes the laptop, and the work stops.

  3. To complete it, the developer keeps a laptop open, and the session cannot be handed over to a colleague. Since 2025, OpenAI and Anthropic have been moving their harnesses to the cloud, and the authors believe their results differ.

  4. Among the challenges the authors list is preserving the agent's state between sessions.

What must be solved during migration

The overall architecture leads to five tasks, each with a real cost

- State. The context accumulated over hours of work must survive a container restart and a network failure. - Isolation. The agent executes code and commands. In the cloud, this requires a sandbox; otherwise, one failed call could affect shared infrastructure. - Access. On a laptop, the agent uses the developer's keys. In the cloud, it needs its own identity with minimal permissions and a defined lifetime for every secret. - Observability.

To understand why an agent deleted a file or used ten times more tokens, you need a log of every step. - Cost. The “model, tool, model” loop

consumes tokens until someone sets a limit.

Without a limit, the bill keeps growing until the first unpleasant surprise at the end of the month.

This set of practices points to what the market calls secure agent access to tools.

The MCP (Model Context Protocol) is important here as the standard way to connect external systems to an agent.

The protocol defines the call format. The company must provide permissions, auditing, and isolation for MCP servers.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

What this means for managers

  1. If your developers already work with agents, you most likely already have shadow infrastructure.

  2. Agents run on personal machines, access repositories with personal tokens, while model costs are charged to corporate cards and are not consolidated into a single budget.

  3. For an IT director, this means three risks: code leakage, irreproducible results, and an uncontrolled budget. The same situation also has an upside.

  4. A cloud agent can be started on a schedule, by an event in a tracker, or when a build fails.

  5. Then it works through the task queue at night while the team sleeps.

  6. This is where TTU, or time to outcome, is measured: the winner is the one who puts the agent on a schedule and tracks its metrics. Whoever leaves the demo on a laptop has only a presentation.

  7. The metrics for a pilot are well established: the share of tasks the agent takes to merge without human edits, and the time from task assignment to merge.

  8. The source contains no figures on cloud harnesses, and citing figures from elsewhere would be fabrication.

  9. These two metrics should be measured in your own pilot before scaling.

How this is addressed technically

We build this setup in stages, in descending order of risk

1. A gateway in front of the model. LLM & Security Gateway sits between agents and providers: Anthropic Claude and OpenAI, as well as GigaChat, YandexGPT, and local Qwen and Llama for closed environments.

The gateway manages token limits, masks personal data and trade secrets, and records an audit trail.

Changing providers does not affect the agents. 2. Tools through MCP.

The agent accesses 1C, Bitrix, Pimcore, or Odoo through separate MCP servers with narrowly scoped permissions. An agent preparing product cards does not need access to accounting. 3. Orchestration.

Event-triggered runs, retries after failures, and human approval for dangerous steps can be orchestrated in n8n or on Kafka when the workload and reliability requirements call for a queue. 4. Development process. In AI-native development, the agent is built into the cycle between “code written” and “working in production”: tests and security checks remain mandatory, and a human performs the review.

Each item has a simple outcome: “the agent completed the task on its own”

, paid for through complex engineering work that is invisible from the outside.

What remains unknown

The authors themselves note that cloud harnesses from major vendors work inconsistently. A hypothesis worth testing is that no universal cloud agent platform will emerge, and companies will assemble a stack of layers in which the model, sandbox, and access layer can be replaced independently. For customers, this is an argument against committing to a single vendor while the market is still unsettled.

Verdict

An agent on a laptop remains a personal tool. It starts delivering value in a company when it has its own identity, budget limit, activity log, and schedule. The cheapest first step is a gateway with limits and auditing: it addresses leakage and overspending risks before you choose which cloud harness to buy.

Discuss the article: Why a Laptop Agent Fails: What the Cloud Harness Provides…

Enter your email or phone number so we can get back to you.

Send via:

By clicking the button, you accept the offer and grant consent to the processing of personal data.

Related solutions

All solutions
Yandex DataLens: CIS's Cloud BI PlatformSaaS Solution: Cloud Services for Business Without InstallationOptimizing Business Solutions for Growth

Articles on the topic

A collection on development management

Related videos

All videos
Shorts "Good Architecture Is Shaped - Around the PERSON" #ktteam #business #IT ArchitectureManagement
Shorts "With the WILL, you are CAPABLE of MUCH MORE than you suspect" #ktteam #business #IT ManagementArchitecture
Shorts The Client MUST Be at the Center #business #ktteam #IT ManagementArchitecture

News on the topic

All news