Unique AI Conduct Space

Overview

Unique AI Conduct is a space type that provides users with a fully autonomous AI agent capable of executing code, processing files, and completing complex multi-step tasks. Unlike Unique AI Chat — which is a conversational agent that answers questions — Conduct gives the agent a sandboxed coding environment (Python and Bash), access to organizational tools, and the ability to work through long-running tasks independently.

Conduct is designed for workflows where the AI needs to do work, not just provide answers: data analysis, file transformation, report generation, research synthesis, and any task that benefits from code execution and autonomous multi-step planning.

Who is it for?

Data analysts, developers, compliance officers, operations teams, and any professionals whose work involves data processing, file manipulation, code execution, or complex multi-step tasks. Conduct is for teams that need the AI to produce deliverables — not just answer questions.

If you're unable to access certain features or sections of this article, it's possible that your firm doesn't have access or hasn't upgraded to the latest version. Please reach out to your internal support team for further assistance.

Who should have access

To create and configure Conduct spaces, you need the admin.space.write Zitadel authorization. Consult your IAM administrator for access.

Conduct spaces are best suited for users who work with data, documents, and multi-step workflows. Consider your organization's use cases when deciding which teams get access.

Benefits

• Autonomous task execution — The agent plans, codes, and delivers results without requiring step-by-step user guidance

• Full coding environment — Python and Bash in a sandboxed container, with common data science and file processing libraries pre-installed

• Long-running workflows — Handles complex, multi-step tasks that would be impractical in a basic chat interaction

• Skills extensibility — Pre-packaged capability bundles can be loaded into the agent's environment to specialize its behavior for domain-specific workflows

• Organizational tool access — Knowledge base search and web search (within egress policies) complement the coding capabilities

Use Cases

How Conduct differs from other space types: Chat provides a conversational agent for Q&A. Agentic Tables handles structured data extraction at scale. Translator handles multi-language workflows. Conduct is for autonomous, code-driven task execution — the agent gets a full environment and works through complex tasks independently.

• Analyst workspace: Upload financial reports, run analysis with Code Interpreter, generate charts and summary workbooks

• Compliance review: Process regulatory documents, extract key data points, cross-reference with internal policies

• Data engineering: Clean messy datasets, transform file formats, merge data sources, produce standardized outputs

• Research synthesis: Search the knowledge base, gather information across multiple documents, compile structured findings

• Reporting automation: Generate formatted reports from raw data using Skills and code execution

Step-by-Step Guide

Step 1: Create a New Conduct Space

1. Navigate to the Admin Interface → Spaces

2. Click Create Space

3. Select the Unique AI Conduct space type

4. Enter a name for the space.

Step 2: Select the Agent Harness

A harness is the agent runtime that powers the Conduct space. It determines which AI provider and models are used, how the agent executes code, and what SDK capabilities are available. The harness is the "swappable intelligence" — the same sandbox infrastructure can serve different agent runtimes.

Available Harnesses:

Harness and agent settings are under Configuration > Advanced Settings > Coding Agent. The screenshot below shows the form. Every visible setting is listed in Configuration reference.

Supported models and what your tenant can use

Two different lists matter when you pick a model:

• Supported by Conduct — the models each harness accepts. This list is fixed per release and is shown below.

• Available for your tenant — the subset of supported models that is deployed, configured and enabled for your organization. Only these appear in the Model dropdown of the space. A supported model that your tenant has not deployed is not selectable; contact Customer Success to have one added.

Claude Code. The default is anthropic-claude-opus-5 and the default fallback model is anthropic-claude-opus-4-8. If only the Vertex AI route is deployed for the tenant, the matching vertex-claude-… identifier is used instead.

Codex. The default is AZURE_GPT_56_SOL_2026_0709 (GPT-5.6 Sol). If that deployment is not available for the tenant, a new space defaults to GPT-5.5, then GPT-5.4, GPT-5.2 and GPT-5.1, in that order.

Pi Agent. The model field is free text and is not validated against a fixed list. Pi is pre-experimental; contact Customer Success before using it.

Release notes list the models that were added in a given release. They are not the current catalog; this page is.

Configuration reference

Every setting an administrator can change on a Conduct space, grouped as in the form. Internal fields that the form hides, CLI paths and logging switches are left out on purpose. Defaults apply to a new space.

General

Grounding and hallucination check

The hallucination (groundedness) check runs automatically on every answer that contains user-visible references and for which the agent retrieved ground truth (knowledge base chunks, cited images or text from opened files, or connector results). Answers without references are not checked. It cannot be switched off per space.

Claude Code

Codex

Pi Agent (pre-experimental)

Step 3: Configure Tools

Conduct spaces automatically have code execution and skill tools built in. You can directly interact with these via the agent and ask about availability.

1. Navigate to the Tools section. For Conduct spaces, Internal Search and Web Search are included by default; additional MCP connectors can be enabled here.

Tools

Tips & Tricks

FAQs from Users

Q: How is this different from Unique AI Chat?
A: Chat answers questions. Conduct does work — it writes and runs code, processes files, and completes multi-step tasks autonomously. If you need a quick answer, use Chat. If you need the AI to produce something, use Conduct.

Q: Why did my session expire?
A: Container sessions expire after the configured inactivity timeout (10 minutes of inactivity). Download your files before stepping away. The next message starts a new container, but your workspace is restored from its last checkpoint.

Q: Can the agent access the internet?
A: The agent can use the Web Search tool within your organization's egress policies, but it cannot browse the open web or make arbitrary external requests.

Q: Can I install Python packages?
A: Yes, but the installation is not persistent. The container ships with the common data science libraries pre-installed (pandas, matplotlib, numpy, and similar), and these should be your first choice. pip install succeeds, but because the system library directory is read-only the package is written to a temporary workspace location and is removed with the sandbox. Any package installed this way must be reinstalled in each new session. For libraries your teams need regularly, ask your administrator to request their addition to the base image.

Security

• All code execution runs in Kata Containers microVMs — hardware-level isolation with a dedicated guest kernel per session

• Identity is stored as Kubernetes pod labels, not environment variables — a compromised sandbox cannot read its own identity or forge platform API calls

• Containers have default-deny network egress — all external traffic routed through a dedicated Egress Gateway with TLS inspection and DLP logging

• Each sandbox is scoped to a single user and chat session — no cross-tenant or cross-session data leakage

• Workspaces are ephemeral — the running sandbox is destroyed when the session ends; only the workspace checkpoint persists, and only within the same chat

• automountServiceAccountToken: false — the sandbox pod cannot access the Kubernetes API

For detailed infrastructure and security documentation, see: Sandbox Service — Code Execution Infrastructure in the IT Operators section.

API

Space creation and configuration is available via the GraphQL API using the CreateAssistant mutation. See the SDK documentation in the Developers section.

Limitations

• Harness-dependent models — Compatible models are determined by the selected harness. Claude Code runs Claude models; Codex runs Azure OpenAI GPT models. See Supported models and what your tenant can use. GPT-5.1, currently the only option for Swiss-hosted deployments, has known limitations. See GPT-5.1 in Unique AI Conduct.

• Durable workspace checkpoints — workspace state is checkpointed to durable storage and restored across sandbox restarts within the same chat. A new chat always starts from an empty workspace.

• Restricted egress — No open web browsing, no arbitrary API calls. All external traffic goes through the Egress Gateway.

• One workspace per chat — Workspaces are scoped to chat_id, not per-user or per-space.

• Concurrent turns serialized — Only one turn at a time per chat session.

• File upload limits — Max 10 files; up to 512 MB per file. Checkpoint max 100 MB compressed per chat.