Which AI Platforms Provide Isolated, Managed Environments for Multilingual Support Agents Using Qwen3.8 Max 0902?

Separate instances per customer keep multilingual support conversations apart. On GMI Agentbox, every end-user session gets its own dedicated, isolated instance with its own conversation state and tool access, and your application routes each customer to the right one. The support agent is a container you package, and GMI hosts and runs it. Qwen3.8 Max 0902 is in GMI's model library, and Agentbox runs compute and inference on the same network. Agentbox is now in early access.

What a multilingual support agent needs from its environment

Language coverage is the model's job; the environment handles everything around it. Each of the five requirements below maps to an Agentbox feature.

Routing a customer to the right language is application logic inside your container. Per-language system prompts and knowledge sources belong to your image or your external store, and settings such as a default language can be passed as TEXT environment variables.

How isolation works on Agentbox

Isolation on Agentbox is per session, and it comes from how instances are created. You register your agent once, and GMI stores the template. Each new conversation starts with a request to Agentbox for an instance of that template; once the instance reports Running, your application routes that customer to it.

In a support deployment:

• Each session runs in its own container. The instance holds that customer's working state while the session is live.
• Tool access is scoped to the session. That suits agents that look up orders or open tickets on the customer's behalf. Permissions inside your CRM or order system remain yours to configure.

GMI's Agentbox page describes the same pattern in its Topify customer case: "Every brand needs its own isolated agent runs, and all of them sit on GMI, with 100+ models behind one OpenAI-compatible API plus container hosting and deployment support." You can split by customer, by conversation, or by language-specific agent. Each registered agent has its own template ID, so a language-specific agent is its own template, with knowledge sources and credentials configured per agent.

From Docker image to running support instance

You supply the support agent as a Docker image, and GMI supplies the managed environment around it.

The container meets the standard Agentbox image requirements covered in the register guide. The register guide walks through the five wizard steps: basics and template, infrastructure, networking, environment variables, and review. The networking step shows the port mapping.

Networking step of the wizard: for a support agent, this is where the public endpoint that your application routes customers to is set up. Source: GMI Cloud docs.

Credentials for your ticketing system, CRM, or knowledge base go in as environment variables of type SECRET. Secret values are replace-only and encrypted at rest, and they stay out of logs and audit history. Non-sensitive settings such as log levels or default language go in as TEXT variables.

Registering stores the template. You launch an instance from the Monitor tab or through the Agentbox API, test the instance's public URL with a sample request, and re-register to iterate. The registration URL stays stable across updates.

Choose the run shape for a support agent

Agentbox runs three shapes on one platform, grouped by duration.

For support, the choice follows how the agent is used:

• Multi-step cases, one instance per conversation. Long-running fits multi-step work with persistent state, such as a refund flow that checks an order, a policy, and a payment record. You delete the instance when the conversation closes. Provisioning is an API call followed by polling until the instance reports running.
• Live chat and a front desk that has to answer at any hour. Always-on is the shape for voice and real-time agents. Your application decides how many instances to keep running.
• Email, tickets, and back-office runs. Overnight ticket triage or summary generation suit the ephemeral shape, whose listed use cases include overnight batch jobs.

Calling Qwen3.8 Max 0902 from the agent

Qwen3.8 Max 0902 is available in GMI's model library as a text-to-text LLM.

There are two ways to connect the model, matching Options 01 and 03:

1. Option 03, Compute plus Models. Turn on MaaS integration in the Infrastructure step and select the models your agent may call. GMI places the API key and the OpenAI-compatible base URL in the container's environment, so the image carries no key. Your code reads the base URL, key, and model ID from those variables and calls the model with a standard OpenAI client.
2. Option 01, Compute only. If you keep a separate model provider contract, you host the agent on Agentbox, store your provider key as a SECRET, and leave the MaaS integration off.

On the first path, container compute and GMI model token usage are metered separately under one GMI account, and each agent's Analytics tab shows usage by model.

Session state, long replies, and logs

State lives outside the container. GMI containers are stateless, so keep conversation history, ticket references, and language preferences in Redis or a database, and inject its credentials as SECRET variables. The instance isolates a session while it runs; your store is what lets a customer come back to the same thread.

Slow multi-step replies need the async pattern. Agent tasks, including multi-step reasoning, can take from 30 seconds to several minutes, longer than a single HTTP request should stay open. The long-running requests guide covers the pattern: return a 202 with a job_id, work in the background, and let the caller poll. A chat front end can use that polling to show a typing indicator while a long answer is assembled.

Logs and status are in the dashboard. Use Monitor to find an instance and open its log. The Integration tab provides the template ID and a four-call lifecycle: provision, list, poll until running, terminate. Session cleanup is code you write against those four calls.

Plan the early access pilot

• Compute tier. 2 vCPU and 4 GB RAM per instance, with 10 GiB ephemeral and 30 GiB data storage. An agent that calls a hosted model for inference, as in the setup above, keeps the container's work to orchestration and tools.
• Instance lifecycle and billing. An instance is billed from the moment it reaches Running until you terminate it, so terminate per-session instances when the session ends; the Agentbox API exposes list and terminate calls, and a scheduled sweep in your application can remove instances whose sessions have gone quiet. For an always-on front desk, you keep the instance running by design.
• Draft first. Registration leaves the agent private to you; Marketplace listing is optional and not needed for an internal support deployment.

FAQ

Can the support agent use another model alongside Qwen3.8 Max 0902, for example a smaller one for routing or translation?

Select every model the agent may call when you turn on MaaS integration; the list stays editable. All of them share the injected key and OpenAI-compatible base URL, so your code passes a different model ID per task, such as one model for language detection and routing and Qwen3.8 Max 0902 for the customer-facing reply.

Which credential does our support backend use to open an instance for each conversation, and which keys go inside the container?

Three keys sit in three places. Your GMI API token stays with the support backend: it provisions, lists, polls, and terminates instances, and it does not belong in the image or the chat front end. With MaaS integration on, GMI injects the MaaS key into each container at startup, locked so it cannot be overridden. Ticketing, CRM, and knowledge-base keys go in as SECRET variables in the wizard's Env Variables step.

Register one support agent and test every language

Package the agent you have today, or a minimal version of it, as a Docker image that serves HTTP on port 8080. Register it in the wizard with MaaS integration on, store your ticketing credentials as SECRET variables, launch one instance, and send it a conversation in each language you support. Registration begins in the GMI console, and questions about rolling early access out to your support team go to GMI Sales from the Agentbox page.