1What it does
A lead Make AI Agent delegates calendar, mail, tasks, projects, contacts, deals, monitoring and memory work to specialist sub-agents. A timer lets Mary work through her own task list and surface what matters.
2Who it is for
- Industry
- Professional services
- Platform
- Make
- Category
- AI Employee
- Tools used
- Make AI AgentsMake AI Sub-AgentsClaude Opus 4.6Claude Sonnet 4.6
3How it works
Every step, in order, is in Workflow at a glance, pinned on the right as you scroll.
- 01 Trigger
Scheduled timer, plus incoming work such as mail.
- 02 AI / action
Run the lead agent on Claude Opus 4.6 for high-level reasoning.
- 03 AI / action
Delegate execution to Claude Sonnet 4.6 sub-agents for calendar, mail, tasks, projects, contacts and deals, Make monitoring, and memory.
- 04 Action
Run regex security checks on input and output at every layer to guard against prompt injection.
- 05 Action
On the timer, work through the agent's task list and surface important work before being asked.
- 06 Action
Handle trivial work autonomously and escalate decisions that need a human. Test nested agent and tool calls rigorously.
- Trigger
- Scheduled timer, plus incoming work such as mail.
- AI step
- The prompts are not published. The recipe below is a starter orchestration pattern, not Mary's original instructions.
4How it went
Make reports daily production use from day one and 2,700+ operations in the past 30 days. Mail, tasks, projects and memory agents each fired hundreds of times autonomously. Marco reports less coordination work and far fewer missed follow-ups. Mary won the community challenge with a 5/5 score from Community Champions.
Only what the source reports. Numbers above come from the source.
5Worth knowing
- Consistent behavior across agents calling sub-agents calling tools was the hardest part. Marco addressed it through rigorous testing.
- One client is planning an implementation. Multi-tenant rollout for small insurance brokerages and expanded write access are roadmap items, with write access dependent on tests.
- The post links a Google Drive demo video. Prompts are not published.
6Recipe
Written by Builds That Work The source did not publish this. We wrote it as a starting point.
Starter orchestration pattern: use an Opus 4.6 lead agent with Sonnet 4.6 sub-agents for calendar, mail, tasks, projects, contacts and deals, Make monitoring, and memory. Check input and output with a regex security gate at every layer. Run on a timer to work through the task list. Starter instruction: "Handle trivial work autonomously. Surface what matters before being asked. Escalate decisions that need a human." Test nested agent and tool calls before expanding write access.