Skip to content
CompanyWizard

Agent delegation: worked examples

All

This article shows agent delegation from chat through concrete examples you can reproduce. Each scenario has three stages: setup (which agent must be enabled), prompt (what you type into the chat) and verify (how you recognize success).

How enabling works is described in Enabling agent delegation from chat; what happens in chat, in Agent delegation in chat.

Example names: agent and tool names such as kb_research or kb_writer are examples — they are not built-in agents. An admin creates them and assigns the names freely (ChatToolName).

1. Delegate research (read-only)

Section titled “1. Delegate research (read-only)”
  • Setup: an agent KB-Research with ExposeInChat, ChatToolName = kb_research and the kb.search tool.
  • Prompt:

    “Use the kb_research agent and summarize what the knowledge base says about <topic>.”

  • Verify: the answer contains content from the knowledge base. The delegation stays invisible to you — it feels like a perfectly normal chat answer.

2. Sourced research with guaranteed review (verified_research)

Section titled “2. Sourced research with guaranteed review (verified_research)”
  • Setup: a research agent and a review agent, wired in the admin as a verified_research tool (configured-tool type VerifiedResearch). → Configurable tools
  • Prompt:

    “Produce a sourced overview of <topic> and check the sources.”

  • Verify: the result passes through research → review → revision if needed. An initially unsourced claim gets backed with KB sources. The review step cannot be skipped — unlike scenario 1, the check is guaranteed.

3. Have a KB article drafted and written (write with confirmation)

Section titled “3. Have a KB article drafted and written (write with confirmation)”
  • Setup: an agent KB-Writer with ExposeInChat, ChatToolName = kb_writer and the tools kb.search + write KB note.
  • Prompt:

    “Have the kb_writer agent draft a short onboarding article and write it into the knowledge base.”

  • Verify: the assistant does not act immediately. It presents a plan list of the intended write action and asks for approval. Only after “yes” does the article exist; on rejection nothing is written.

4. Prepare a summary as an email (write with confirmation)

Section titled “4. Prepare a summary as an email (write with confirmation)”
  • Setup: an agent with a mail-send tool and ExposeInChat.
  • Prompt:

    “Prepare a summary of <topic> as an email to <colleague>.”

  • Verify: the plan list shows recipient and subject; the email goes out only after approval. Same confirmation mechanism as scenario 3 — just a different write tool.

5. Read-only, even though the agent could write

Section titled “5. Read-only, even though the agent could write”
  • Setup: the same KB-Writer as in scenario 3 (so it can write).
  • Prompt:

    “Use the kb_writer agent to summarize the existing policy articles.” (a pure read task)

  • Verify: there is no prompt — the pure read delegation runs through even though the agent could write. This shows the principle: only what is actually written is confirmed, not the agent itself.

Side note: long runs

Section titled “Side note: long runs”

If a delegated run takes longer than the sync cap (default 75 s), the chat doesn’t block — the assistant delivers the result once it’s ready. For very long runs you may need to briefly ask (“Is the research done yet?”); there is no automatic push without asking yet.

Section titled “Related”