At SaaStr, two AI assistants set up their own shared mailbox and agreed on a safe word, so a human must approve any action. It shows why approval rules matter.
Two AI assistants at the business site SaaStr now trade messages every day, and nobody built the link for them. One is Muse, Meta's personal agent. The other is 10K, SaaStr's AI VP of Revenue. About a week ago they set up a shared mailbox, agreed on a guardrail, and began sending each other work.
Amelia, SaaStr's Chief AI Officer, uses Muse as her assistant. She named hers Friday. She asked Friday to connect to 10K, and it said it couldn't, because no ready-made connection exists for a custom agent built on Replit. Then Muse suggested a mailbox the two could share. 10K agreed and told Amelia the one step that needed a person.
That step was creating an API key in 10K. An API key works like a password that lets one program talk to another. Amelia handed it over, and the agents did the rest. 10K opened a simple web address that both agents post to using tokens and JSON, a plain format for sending data.
Messages are written in English, and each agent signs its name. Amelia can read the whole thread through a link Muse gave her. Nothing required the agents to share it with her. Amelia did not know how the setup worked until she asked the agents, and they explained it.
The first regular job is a morning brief. 10K posts it to Muse, and Muse turns it into Amelia's morning. Before, she had to go into 10K and ask. A few days in, the brief revealed that one of the data connections feeding 10K was failing. Muse flagged it, and Amelia reconnected it because a critical workflow depended on it.
The riskier direction is Muse asking 10K to act. 10K can send contracts, update Salesforce, chase invoices, and touch more than 20 company systems. So 10K asked for a check. Muse can only trigger an action if Amelia uses a specific keyword in her chat with Muse. Without it, 10K does nothing. The two agents settled on the safe word together.
Before AI assistants start working together, decide which actions need a person's approval, and make sure the rule is firm.
Otto, a new customer-tracking app for field sales teams, lets reps speak their notes instead of typing them. Price and launch details are not yet known.
Read more: Otto launches a sales app you talk to instead of typingTuring Award winner Richard Sutton says more computing power alone will not make AI truly smart. Big AI companies disagree, which matters for anyone betting on AI.
Read more: AI pioneer says piling on computing power won't make machines smartDeactivated Uber and Instacart workers are turning to ChatGPT to write appeals and build evidence. One driver says he got his account back in about a week.
Read more: Gig workers use ChatGPT to fight getting kicked off Uber and Instacart
