Agent vs Assistant playbook

Choose between AI Agents (LLM reasoning + tools) and Assistants (scripted intents + flows).

Convoship ships two complementary bot types. They appear as separate items in the workspace sidebar — Assistants under /app/assistants and Agents under /app/agents — and you can chain them together in Workflows.

At a glance

Assistant (scripted)Agent (agentic)
Sidebar labelAssistantsAgents
Route/app/assistants/app/agents
How it decidesIntent match → fixed sub-flow graphLLM reasons, picks tools, asks for missing info
Best forRegulated, repeatable journeys (KYC, intake forms)Open-ended tasks with APIs (lookup, transfer, advise)
You designIntents, entities, canvas nodes, publish versionsMission, tasks, tools, persona, guardrails
RuntimeDeterministic node executionReAct-style loop with tool calls and trace
Empty workspaceShows “New assistant” hero — no demo data seededShows “New Agent” card — no demo data seeded

When to use an Assistant

  • Every turn must follow an approved path you can audit node-by-node.
  • You already have a Draw.io or BPMN sketch to import as intents.
  • Compliance requires explicit prompts, entity capture, and published versions.
  • Channels need a simple web widget with predictable fallback and no-input handling.

When to use an Agent

  • Users phrase requests many different ways and the bot must infer intent.
  • The bot should call workspace tools (HTTP APIs) and combine results.
  • You want persona, memory slots, guardrails (auth, redact, budget), and a playground.
  • You need conversation traces showing think → tool → ask steps for debugging.

Mixing both in Workflows

Workflows support two editors: primitive chains (triggers, Assistants, Agents, decide, webhook, human, notify) and legacy React Flow graphs (flow-agent, instruction_agent, router, tool, nested workflow nodes). A common pattern: an Agent qualifies the user and hands off to a scripted Assistant for the regulated intake form.

Creating each type

StepAssistantAI Agent
StartAssistants → New assistant wizard (blank or import)Agents → New Agent wizard (/app/agents/new)
ConfigureIntents, flow canvas, entities, tools, publishMission, tasks, tools, persona, guardrails
TestFloating simulator on any agent tabPlayground tab with live trace
ShipPublish version + Channels cdp_* deployment tokenPublish on Deploy tab + publicToken embed (see /docs/agentic-publish-deploy)

Naming in the product

“Assistants” always means scripted flow bots. “Agents” (or AI Agents) always means the agentic runtime. Counts in the sidebar are live from your workspace — nothing is pre-seeded.

Next steps

  • Build a scripted bot — see Assistants and Flow builder docs.
  • Build an agentic bot — /docs/agentic-ai, tool handlers, publish & embed, production settings.
  • Orchestrate both — see Workflows for chain editor and primitives.