Orchestrator & Agents

In Beyond Better, the orchestrator plans the work and agents do parts of it in the background. Their reports come back to the orchestrator, which brings everything together for you.

The Idea in One Minute

  • The orchestrator is the assistant you talk to. It understands your objective, decides how to split the work, and answers you.
  • Agents each take one task, such as "summarise these five reports" or "check the pricing pages of these competitors". They work in the background, in their own conversation, and send a report when they finish.
  • Personas decide how each of them works: the LLM, instructions, tools and limits. You don't steer agents one by one; you shape them through personas.

Why it helps: independent parts of a job run at the same time, each agent starts with a clean, focused context instead of the whole conversation, and the orchestrator's own conversation stays short enough to keep the big picture in view.

How It Works

  1. Plan: the orchestrator works out which parts of your request can be handed off.
  2. Start agents: it writes a task for each agent, with the background and instructions the agent needs, and picks a persona for each one. An agent can't see your conversation, only its task.
  3. Carry on: the orchestrator doesn't wait. It can keep working, answer you, or finish its turn while the agents work.
  4. Reports arrive: as each agent finishes, its report appears in the conversation. If the orchestrator had finished its turn, it picks the report up and continues on its own.
  5. Bring it together: the orchestrator combines the reports into the answer you asked for.

How many at once: up to 5 agents run at the same time in a conversation. Any more wait in a queue and start as others finish. See How many agents run at once to lower or raise this.

Agents starting agents: if a persona allows it, an agent can split its own task and start agents of its own, up to 3 levels deep. By default, agents can't.

Always available: starting and managing agents doesn't depend on which tool sets you select, so turning tool sets off doesn't stop the orchestrator from delegating. Use a persona's Delegation depth to stop it.

Asking for Agents

The orchestrator decides for itself when agents are worth it. Simple requests, and work that needs a lot of back-and-forth with you, are usually done directly. You can also ask:

"Use agents to review each of these three research papers, then compare their methods."

"Start a Researcher agent for the background reading while we work on the outline."

What You See While Agents Work

Agents started together appear in the conversation as an agent group. For each agent it shows the task, the persona, a status, the time it has been running, and its latest progress. Agents started by an agent appear as a group nested inside it.

Three agents working in parallel: Marketing plan vs sales, Menu vs survey and Wholesale price change notes, each Running with its persona, who started it, elapsed time, tool calls and latest progress, and Message and Cancel buttons
Agents working in parallel, each in its own group
StatusWhat it means
QueuedWaiting for a free place among the agents that can run at once.
RunningWorking on its task.
WaitingPaused for a person: a question, an approval, or an MCP server sign-in. A parent agent shows "Waiting for its agents".
StoppingFinishing its current step after a cancel.
CompletedFinished; its report is in the conversation.
Needs more workFinished, but reported that the task needs more work.
FailedCouldn't finish, for example because a data source its task uses was removed, or BB restarted.
CancelledStopped by a person or by the orchestrator.
EndedNo longer running, but its outcome isn't known in this view yet. The agent's report says how it finished.

The status bar above the message box says in plain words what is happening, such as "Editing report-draft.md" or "Starting 3 agents", with a line for each running agent.

Status bar reading Waiting for agents to report, with a line for each agent: Wholesale price change notes reading two files, Menu vs survey and Marketing plan vs sales thinking
The status bar while three agents work

Steering Running Agents

  • Message: send an agent extra instructions. It reads your message after its current step. Everyone in the conversation sees the message, as yours.
  • Cancel: stop one agent. The answer and other agents carry on. Cancelling an agent also cancels any agents it started. If someone else started it, BB asks you to confirm first.
  • Stop: the Stop button's menu can also cancel running agents, or halt everything. See Managing Collaborations.
  • Answer questions: agents can ask a question and wait for an answer. Anyone who can write to the conversation can answer it. See Questions & Approvals.

Message and Cancel appear on queued, running and waiting agents, for anyone who can write to the conversation. The orchestrator has the same abilities: it can message, cancel and wait for its agents, and look back through an agent's work.

If BB restarts: agents that were running or queued stop and show as Failed ("It was stopped by a server restart."). Their reports still reach the conversation, so the orchestrator knows what was done and can start them again.

Whose Access Agents Use

Normally an agent works with the access of the person whose message started it. In a shared team workspace, your team can set up agent accounts so agents run with an account of their own. The agent group then shows a label such as "As Research agent, started by Sam".

Time and Cost

  • Agents finish large jobs sooner because they work at the same time, but each one also costs credits: it starts with its own instructions and does its own reading.
  • The agent persona's model matters most for cost, because many agents may run. A faster, cheaper LLM is often enough for reading and summarising.
  • Use Agents at once and Delegation depth in a persona's limits to keep spending bounded.

Example Uses

Literature review

  • Agent: urban planning studies
  • Agent: agricultural adaptation studies
  • Agent: economic impact assessments
  • Orchestrator: one combined review

Multi-channel content

  • Agent: blog plan
  • Agent: social media calendar
  • Agent: newsletter outline
  • Orchestrator: consistent messaging and timing

Customer analysis

  • Agent: purchase history
  • Agent: support conversations
  • Agent: website behaviour
  • Orchestrator: customer profiles

Software project

  • Agent: API design
  • Agent: data transformation
  • Agent: error handling
  • Orchestrator: make the parts fit together

Getting Good Results

  • Say what the finished result should look like. The orchestrator passes that on to its agents.
  • Give agent personas clear names and descriptions, so the orchestrator picks the right one for each task.
  • Keep agents' tool sets as narrow as their work allows, and use "Ask a person to approve each one" where you want a say.
  • Older guidelines that mention delegate_tasks still work, but the tool is now called start_agents.

Next Steps

Last updated: November 16, 2026