Roles and hand-offs

Build-along · ~18 min at your own keyboard

More agents is not better by itself. Every extra role is another set of instructions to keep, another hand-off that can break, and more cost. Split work into separate roles only when there's a reason: the step needs different information or permissions, it's easier to test on its own, or a check is more honest when done by something that didn't write the draft. Some steps don't need AI at all — a filter rule or a spreadsheet formula is cheaper and more reliable. A common shape is intake (sorts what came in), drafter (makes the thing) and reviewer (checks it against your rules), but one well-limited agent plus your review can be the right design. Whatever you choose, each role has one job, one input, one output and one hand-off, kept as plain text you can read.

Ladder · Level 6The long version: Roles and hand-offs ↗

Do this

0/7

Checkpoints

▸Each step is marked as its own role, combined, or not AI, with a reason.
▸Each AI role has a one-sentence job and each hand-off has a written format.
▸One real item passed through every role.

The three ways this goes wrong

The Drafter re-sorts the inbox or the Reviewer rewrites the draft. — Add 'You do not do X' to that role. The Reviewer flags, it does not rewrite. If two roles keep doing the same work, merge them.
Hand-offs arrive in a different shape each run. — Give the format as a template with field names and an example. Say 'use exactly this shape'.
The Reviewer approves everything. — Give it your never-do list as a checklist and require it to quote the line it checked against.

Dated facts in this lesson: ChatGPT has agent, Projects (project-only memory, files, instructions), memory toggles in Settings > Personalization, and Scheduled Tasks at chatgpt.com/schedules. Workspace agents ("ChatGPT Work", July 2026) are on Business and Enterprise. Agent quotas: Plus 40 messages a month, Pro 400. Secondary source (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Claude Cowork has persistent sessions, memory, schedules and Chrome. It needs Pro ($20/month) or above (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Gemini Spark is a 24/7 agent with Skills and Schedules. Schedules need Pro or Ultra, personal accounts only, up to 50 schedules. A Mac version exists. Spark is a beta (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Microsoft 365 Copilot has Notebooks, Agent Builder agents, Pages, Copilot Cowork and Researcher and Analyst agents. Personal Copilot has Memory and Personalization settings, and Notebooks with Microsoft 365 Personal, Family or Premium. Secondary source (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Grok has Projects (files and instructions), "Personalize with Memories" under Data Controls, and Automations on a paid tier. SuperGrok Heavy ($300) runs parallel agents. Secondary source (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it); Perplexity has Memory (also on Free), Spaces (files and instructions; Pro allows 50 files per Space; not on Free), Tasks (at most hourly) and Computer on Max. Secondary source (researched Oct 4, 2026, not yet confirmed by a person: check the source before you rely on it).

Share what happened

Paste your Who does what table with one reason for each role decision, the hand-off formats, and one real hand-off from a test run.

Strip API keys, passwords, and tokens before you paste. Swap them for [KEY].

🔒 Don't paste confidential, personal-health, or payment info.