When should I split one agent into two?

When its prompt starts doing two jobs — e.g., 'enrich the row AND write the email.' Separate prompts beat one crowded one.

Do the agents talk to each other?

No — they share an artifact. Agent A writes rows to the sheet; Agent B reads rows from it. The sheet is the contract.

Does chaining need Pro?

If the pipe is a connector (Sheets, a queue) yes; two paste-and-run agents chained by hand stay free.

Multi-bot workflows — chain agents through shared artifacts

Split work across agents by giving each one job and one shared artifact — a sheet or doc is the pipe between them.

Steps

Name the artifact between the two jobs — usually a sheet with defined columns, or a doc with a fixed section shape.
Write the upstream agent so its output IS the artifact’s format: exact columns, exact section names.
Write the downstream agent to read only that artifact — its prompt assumes the format exists, nothing else.
Chain them manually first: run A, inspect the artifact, run B. When the handoff is clean, schedule A before B with a gap.

The collection pattern

This is exactly how our role collections are designed — several collections chain templates through a shared artifact — sales is the canonical one: the enricher fills the sheet, the outbound drafter writes from it, and the follow-up drafter reuses the same account context.

What breaks chains

Format drift — A writes “company size” one week, “size” the next. Pin the column names in A’s memory AND B’s prompt.
Skipping the artifact inspection — a chain is only as honest as the middle file. Look at it weekly.
One mega-agent instead — tempting until the prompt hits context limits and does both jobs badly.

Next: see the chains

Browse the role collections — each workflow section is a multi-bot chain in the wild.