Skip to main content

Manager History

Manager History (shown as Multi-Agent History in the builder) is the review surface for your multi-agent workflows. It shows every orchestration run — which sub-agents the manager delegated to, what task it gave each one, how each delegation went, and the full nested transcript of every sub-agent conversation.

What it's for​

Multi-agent runs have more moving parts than a single agent conversation: the manager reasons about the request, picks sub-agents, phrases a task for each, and synthesizes their results — possibly over several iterations. Manager History lets you:

  • Audit delegation decisions — see which agent got which subtask and why the manager's reasoning led there.
  • Debug failures at the right layer — distinguish a manager that delegated poorly from a sub-agent that executed poorly, by drilling from the orchestration view into each sub-agent's own transcript and tool calls.
  • Track quality per version — the same monthly stats and version performance breakdown as agent history, so you can compare published manager versions.

From the dashboard sidebar, open Agentic Workflows → Multi-Agent History. Pick a multi-agent in the Select Multi-Agent dropdown, set the date range (defaults to the current month), and optionally include your own sandbox conversations.

Key concepts​

Run rows​

Each conversation row shows its evaluation status, close date, manager version, message count, Agents Used (how many sub-agent invocations occurred), Iterations (how many orchestration cycles the manager ran), and the conversation ID.

The orchestration viewer​

View Details opens the manager conversation viewer. Alongside the user's inputs and the manager's replies and reasoning, each delegation appears as an Agent Invocation card showing:

  • the sub-agent's name and the exact task description the manager gave it,
  • the delegation status (complete, partial, or failed),
  • request details — the variables passed to the sub-agent,
  • the sub-agent's response summary and its own conversation ID,
  • an expandable Agent Conversation Context — the sub-agent's full nested transcript, including its tool requests and responses.

An evaluation side panel mirrors the one in Agent History: automated scores plus human evaluations that take precedence.

Billing note: a manager conversation is billed based on the number of sub-agents configured on the manager — the sub-agent invocations inside the run are part of that one manager conversation, not separately billed agent conversations.

How to debug a bad multi-agent run​

  1. Open Agentic Workflows → Multi-Agent History, select the manager, and find the run (unsuccessful runs are marked with a red icon).
  2. Click View Details and read the manager's reasoning before each delegation — was the right sub-agent chosen, and was the task phrased correctly? If not, the fix is in the manager's instructions or the sub-agents' descriptions.
  3. If the delegation looked right, expand the invocation's Agent Conversation Context and inspect the sub-agent's own transcript and tool calls. A failure there is fixed on the sub-agent (instructions, tools, or arguments).
  4. Record a human evaluation with notes so the run counts correctly in the stats.