Multi-Agent Orchestration: Let One Agent Lead the Others

BotTrottr Team · · Product

One agent doing everything is a fine start. But the moment a job has distinct parts — research a topic, draft the summary, check the numbers — a single prompt stuffed with every instruction starts to creak. Today we’re shipping multi-agent orchestration: an agent can now act as a coordinator, delegate pieces of work to other agents in its swarm, and stitch their answers into one result.

How it works

Turn on Orchestration when you deploy an agent and it gains one new tool, delegateToAgent. When the agent decides a sub-task belongs to a specialist, it calls the tool with the peer’s ID and a complete set of instructions. BotTrottr runs that peer exactly as it would run it on its own — same model, same system prompt, same connectors, same skills — and returns the peer’s text answer to the coordinator, which carries on and writes the final response.

A few things are worth knowing:

  • Delegates start fresh. A delegated agent doesn’t see the coordinator’s conversation. The coordinator has to put everything the peer needs in the task, and the coordinator’s prompt tells it so.
  • Answers come back as text. The coordinator is told to expect plain text results and to combine them itself.
  • Each agent keeps its own access. A peer uses only its own credentials and connectors. Delegation doesn’t let one agent borrow another’s keys.

Same swarm only

Agents can only delegate to agents in the same swarm — and only to agents you own. Assign the coordinator and its specialists to one swarm on the Deploy Agent card, then open the new collapsible Orchestration section:

  1. Toggle This agent can delegate to other agents.
  2. Pick a roster of swarm-mates, or leave every box unchecked to allow the whole swarm.
  3. Optionally allow the agent to spawn a copy of itself for parallel work.

The check is repeated on every run. If you later move an agent out of the swarm, delegation quietly switches off.

A shared data folder for the swarm

Passing a paragraph of text back and forth works for small answers; bigger hand-offs need a place to put files. Every agent in a swarm now has access to a shared data folder alongside its private storage. The four storage tools — list, read, write, delete — accept a shared option that points them at the swarm folder instead of the agent’s own.

A research agent can write findings.md to the shared folder, and the writer agent can read it, without anything being pasted through a prompt. The folder is scoped to you and that swarm, so other accounts’ agents never see it.

Guard rails

Every delegation is a full model run, so cost and runaway behavior matter. These limits are built in:

LimitValue
DepthOne level — a delegated agent can’t delegate again
Delegations per run6
Running at the same time3
Time per delegated run2 minutes
Result size passed back8,000 characters

Delegated runs are recorded in the peer agent’s run history, tagged with who delegated them, and their token usage is billed like any other run — so nothing is hidden from your usage totals. Offline agents are left off the roster.

Plans

Orchestration is available on BYOK and paid plans. Free-plan agents run on BotTrottr’s included key, and a fan-out of model calls multiplies that spend, so it isn’t offered there. The same rule is enforced on the server, not just in the form.

Good first uses

  • Specialist routing. A coordinator with a research agent, a writing agent, and a reviewer, each with its own prompt and tools.
  • Fan-out. Allow “copy of itself” and have a coordinator split a long list into independent chunks.
  • Escalation. A fast, cheap coordinator that hands only the hard questions to an agent running a stronger model.

What’s next

Today the Orchestration section lives on the Deploy Agent card and applies at creation. We’re looking at editing it on existing agents, delegating across the members of a team swarm, and richer hand-offs than plain text. If you try it, tell us what you’d want it to do next.

← Back to all posts