Inside the AI Engine

How Conduit Agents Work explains the four system components: Inbox, Build Surfaces, Workflows, and Dashboards. This page goes one level deeper and follows a single message through the engine, showing how the pieces you configure in Build - Knowledge, Skills, Tools, Persona, Guardrails, and Behavior - actually get used at response time.

​

The Response Flow

Every reply starts the same way: something triggers the agent, and the agent works through its configuration in order before a message goes out.

​

Retrieval: Where Facts Come From

Knowledge is organized as a folder tree of pages, folders, and source-backed files. When a message comes in, the agent isn’t searching your entire workspace blindly - it’s pulling in the pages relevant to the question, including any Knowledge that a matched Skill references directly. Composable Knowledge keeps this retrieval clean: a global page (like a cancellation policy) stays the source of truth, and scoped pages reference it and add only local exceptions, instead of duplicating the whole policy per listing or product. That means retrieval surfaces one coherent answer instead of several conflicting copies of the same fact. Source-backed Knowledge (documents, links, synced integrations) goes through the same retrieval path once it’s indexed. Content still in a Syncing or Waiting on quota state isn’t available to retrieve yet - see Sync Status.

​

Matching The Right Skill

Skills are conditional: they apply when the current scenario matches their trigger, not on every message. An agent only draws from the Skills assigned to it in its Skills tab, so narrowing what’s assigned to an agent is part of narrowing what it can match against. A good Skill is specific enough to trigger in the right scenario and narrow enough that it doesn’t compete with unrelated Skills - see Maintenance. When a Skill’s scenario depends on a reusable sub-process (escalation, refunds, rescheduling), it references that Skill by name rather than re-describing it, and the agent follows the chain at match time.

​

Calling Tools Inside A Skill

A matched Skill tells the agent when to act, and the assigned Tools tell it how. Each tool’s description is part of the instruction surface the agent reads at call time - what the tool does, when to use it, what information to collect first, and what to do if the call fails or succeeds. That’s why tool descriptions in Conduit are written for the model, not just for a human reading a settings page. Tool access is scoped per agent: granting a tool to one agent doesn’t grant it to another, and Operator has its own separate tool settings from any customer-facing agent. Keeping tool access narrow keeps this step focused - an agent performs better calling from a short, relevant list of tools than from every tool available in the workspace.

​

Guardrails And Behavior Before Sending

Guardrails and Behavior do different jobs at this stage:

• Guardrails are hard boundaries, checked against every response regardless of which Skill or tool produced it.
• Persona shapes tone, role, and formatting so the reply sounds like your agent.
• Behavior controls the deployment mechanics around the reply - escalation rules, working hours, response delay, structured output, and hooks that can call tools, HTTP endpoints, LLM handlers, or other agents before and after a run.

Escalation rules decide which messages get handed to a human instead of answered automatically, and Hours sets the agent’s working hours - see Behavior Settings.

​

Why This Is Transparent

Because retrieval, Skill matching, tool calls, and Guardrail checks are discrete steps rather than one opaque pass, Conduit can show you which Knowledge pages it pulled in, which Skill it matched, and which tools it called for any given reply. That’s what “transparent, showing reasoning and sources” means in practice: you can inspect the path a response took, and approve, edit, or teach the agent directly from the Inbox when it should have taken a different one.