Documentation: integration, capabilities, and limits
Hub for building Duale AI integrations: Python SDK, Agent harness, Tools, Libraries, identity, security, and production limits.
This guide covers core Task workflow, identity and access, integration capabilities, production boundaries, and focused troubleshooting entry points.
- Core Task workflow: Python SDK, Agent harness, outcomes, and context continuity.
- Identity and access: person, Agent Identity, credential, group, and system origin.
- Integration capabilities: Tools, Libraries, Web research, Code Execution, and model routing.
- Production boundary: security review and scope limits.
- Focused entry points: SDK troubleshooting, errors, content marking, and Documentation MCP server.
Summaries were generated by AI.
Start with the Python SDK to submit a Task request and receive its typed result. Then use the Agent harness to understand how a Task uses Tools, reaches an outcome, and supplies history for an independent continuation.
When a page directs you to the Dashboard, use the URL from your deployment handoff. The managed service uses the Duale AI Dashboard.
Start here
Build the core integration and Task workflow in this order:
- Python SDK—install the Public Preview package, submit typed Task requests, host custom SDK Tools, and attach documents.
- Agent harness—connect Task inputs, Tool use, outcomes, and permission for external actions.
- Task outcomes, deadlines, and Stop—follow a Task’s work and know what its deadline and Stop request mean.
- Context continuity—continue from selected history, compare independent branches, and understand retained information.
Configure identity and access
Use Identity to distinguish a person, Agent Identity, credential, group, and audit-only system origin. Follow its links to policies, credential lifecycle, and IAM limits, or Library access for one Library’s Grants.
Add the capabilities your integration uses
- Tools—choose between custom SDK Tools and organization-managed MCP Connections, then review assignment, availability, and call evidence.
- Libraries—give Agent Identities governed document access and make bounded retrieval Tools available during execution.
- Web research—learn how a Task finds and reads public pages, and what source access does not prove.
- Code Execution—understand the bounded built-in Tool for running agent-written Python or JavaScript.
- Model routing strategy—design Tenant model pools, regional boundaries, endpoint redundancy, and model upgrades; document the endpoints, private connectivity, and network paths each environment needs.
Review the production boundary
- Security—review shared responsibility, data protection, the threat model, and available audit evidence.
- Scope and limits—review intended use, operational limits, records, and customer duties before production.
Use a focused entry point
- If a Task, Tool, or document operation fails, start from the symptom in SDK troubleshooting; Errors and reliability lists typed failures, retries, and timeouts.
- If you need to verify whether an answer is marked as machine-generated, use Content marking.
- If an AI client needs to search and cite these pages, connect the read-only Documentation MCP server.