Agents and workflows

DataQI employs an autonomous agent architecture that breaks down complex enterprise requests (e.g., “Get me the latest support tickets AND the current CRM logs”) into actionable steps.

Agent classifications

Within DataQI, agents are categorised by their function and accessibility:

• Main agents
  • Main agents: Maintain the high-level workflow and interact with the user.
  • Sub-agents: Dispatched in the background to handle specific, isolated tasks.
• Accessibility
  • Chattable agents: Configured to appear in the user interface for direct conversation.
  • Non-chattable agents: Operate strictly in the background as headless workers.
• Origin
  • System agents: Built-in core agents that handle platform-level routing and operations.
  • Customer-specific agents: Custom agents tailored to your enterprise’s unique environments and tools.

Skills, tools, and workflows

To complete requests, agents rely on three key parts:

• Skills: Domain-specific instructions and boundaries that define an agent’s persona and scope of responsibility.
• Tools: Actionable capabilities granted to an agent (such as MCP connections or internal API calls) that allow it to fetch data or mutate state.
• Workflows: Multi-step tasks where agents figure out what to do, plan their approach, and use tools in order to reach a goal.

Parallel delegation

Rather than forcing you to break your request down into sequential steps, the main agent identifies what information is needed and assigns tasks to specialist agents at the same time.

graph TD
    User([User Request]) --> Orch(Orchestrating Agent)
    Orch -->|Delegates Task 1| SA1[Support Sub-agent]
    Orch -->|Delegates Task 2| SA2[CRM Sub-agent]
    SA1 -->|Fetches Tickets| D1[(Service Desk)]
    SA2 -->|Fetches Logs| D2[(CRM System)]
    D1 --> SA1
    D2 --> SA2
    SA1 -.->|Results| Orch
    SA2 -.->|Results| Orch
    Orch --> Output([Unified Response])

Once the sub-agents complete their work, the orchestrating agent collates the results and synthesizes them into a single, unified response. This fan-out approach significantly reduces response times for complex workflows.

Partial failure resilience

If one sub-agent encounters an error or times out while working, the entire request does not fail. DataQI gracefully collates the successful results alongside a structured error report, informing you exactly which parts of your query succeeded and which parts require attention.

Security: trust partitioning

Agents often combine your natural language queries with external data (like documents or webpage context) to complete tasks. To defend against prompt injection from untrusted data, DataQI employs a strict Trust Partitioning mechanism. For more details on how DataQI protects agent instructions, see the Security boundaries guide.