#ai#management#automation

Agents in Production: Cheaper Than Control

Claude in GovCloud, MCP promotion for agents, and multi-agent evaluation reveal AI’s real production cost.

  • The occasion: five AWS publications in one week
  • What happened
  • Where companies lose the outcome
  • Migration: manual work prone to errors

The occasion: five AWS publications in one week

In one week, AWS published five pieces about operating AI agents. Claude Opus 5.5 and Sonnet 5.5 became available in AWS GovCloud (US) regions for tasks with regulatory requirements, including ITAR. Amazon Quick gained automated agent migration between accounts. Explainability assessment became available for multi-agent systems. Based on our project experience, model selection takes about a day, while building the control layer around it takes a quarter, and the outcome depends on that layer. This is the idea we examine.

What happened

AWS examined five scenarios: - Claude Code on Amazon Bedrock in GovCloud: AI development for organizations subject to export and regulatory controls. -

The `aws-ai-ml` skill in Agent Toolkit for AWS.

It gives coding agents (Kiro, Claude Code, Codex) knowledge about inference optimization in SageMaker: the agent benchmarks endpoints, compares runs, and generates code using SageMaker Python SDK v3. -

An idempotent MCP server on Bedrock AgentCore that moves agents, connectors, knowledge bases, and Amazon Quick flows from a dev account to production. -

Evaluating multi-agent systems for usefulness and explainability

- GLM 5.3 by Z.ai: a mixture-of-experts model with 753 billion parameters for coding and long-running agent tasks.

It is available on Bedrock without dedicated inference infrastructure.

We see these announcements as one theme: AWS is describing how to operate agents and releasing tools for that purpose alongside models.

Where companies lose the outcome

A typical scenario looks like this

A team builds an agent in two weeks, it performs well in a demo, and the manager approves the launch. Then three things become clear.

An agent cannot be moved to production without manual reconstruction.

No one can explain why it made a particular decision.

The security team prohibits sending customer data through an external API.

The budget is spent, the metric has not improved, the project is labeled an “experiment,” and the manager who approved it is held responsible for the wasted money.

AWS materials address these three gaps: migration between environments, verifiable quality, and an acceptable data perimeter.

Migration: manual work prone to errors

  1. AWS explicitly describes moving Amazon Quick resources from development to production as a manual, error-prone task.

  2. An agent consists of several parts: the agent itself, action connectors, a knowledge base, flows, and spaces.

  3. If some components are moved and others are not, production will contain an agent that calls a nonexistent connector. The solution is built on an MCP server.

  4. Idempotency means that rerunning an operation produces the same state without duplicates.

  5. Auditing means that each deployment leaves a record of who deployed what and when.

  6. Engineers have long applied this discipline to releases, and it is now being extended to a new type of artifact.

  7. The MCP protocol acts as a standard seam between an agent and a company's systems: the agent uses it to access ERP, PIM, or CRM, while administrators use it to manage the agent's lifecycle.

Assess where AI can deliver impact in your process

clients@kt.team Telegram @kt_team_it

Evaluation: how to prove an agent is useful

  1. Multi-agent systems coordinate several specialized agents and make decisions in a chain: supply planning, financial analysis, and customer service.

  2. An error made by the first agent is passed to the second, and the final answer looks convincing even when the chain broke midway.

  3. That is why AWS proposes evaluating two parameters: the usefulness of the result and the explainability of the path to it.

  4. The practical takeaway for a company is this: if you cannot reconstruct which agent decided what and based on which data, you cannot improve the system or be accountable for it to regulators and customers.

  5. A quality metric must exist before launch.

  6. Without it, there is no way to confirm either success or failure after launch.

Perimeter: regulated data and model selection

The availability of Claude in GovCloud shows how the question has changed in regulated industries.

Companies used to ask whether they could use AI.

Now they ask within which perimeter

AWS includes a caveat with the publication: the approach is not suitable for every organization and must be checked against its own compliance requirements.

This caveat should be read literally

GLM 5.3 on Bedrock adds a second aspect: an open model can now be obtained as a managed service.

The same task can be solved with multiple models at different prices and under different jurisdictions.

This requires a unified layer through which traffic to models passes: access policies, data masking, request logging, and routing. In our practice, this layer is called LLM & Security Gateway.

It allows you to change the model without rewriting applications.

Development speed: a skill instead of a prompt

The `aws-ai-ml` skill shows how an engineer’s work is changing

The agent receives specialized domain knowledge, runs benchmarks independently, compares configurations, and produces executable code. This is the principle of AI-native development: we measure value by the time from task to a working result (TTU).

The amount of generated code says nothing about this evaluation. The “select inference configuration” operation

looks simple, but requires knowledge that used to reside with one or two specialists. That knowledge can now be packaged and reproduced.

What to check before launch

Before deploying an agent to production, you should get answers to four questions: - Deployment reproducibility.

Can the agent be deployed to production in one idempotent, logged operation? - Quality metric.

Has a measurable criterion for usefulness and explainability been defined before launch? - Data perimeter.

What data is sent to the model, where does it run, and who can see it? - Model replacement.

How much effort will it take to switch to another model if the price or terms change?

If even one question has no answer, the company has a demo, not a system.

Conclusion

The five AWS publications read like an operations manual for agents.

The model takes up one line; the rest describes releases, auditing, evaluation, and the data perimeter.

These elements distinguish AI that moves the metric from AI that looks good in a presentation.

The business pays for the outcome, and the control layer around the model delivers it.

Engineers build it in advance, before the first launch.

Discuss the article: Agents in Production: The Model Costs Less Than…

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 Cases

All cases

Related solutions

All solutions
AI monitoring is automatic control of processes and eventsAI assistant for Rosreestr: EGRN extracts, verification, and monitoringGEO Website Optimization for AI Search and LLMsAI Accountant in Your Environment: Implementation Cost

Articles on the topic

AI in Business Collection

Related videos

All videos
Shorts What “KPI” did Amazon set for developers in the AI era? AITeams & hiringManagement
Shorts Can't Keep Up with AI? Build Communication and Accountability AI
Shorts AI does not empower the weak! AI

News on the topic

All news