SaaStr's AI Agents Built Their Own Channel — and a Safe Word

Two AI agents at SaaStr set up their own communication channel in roughly an hour, using a single human-generated API key — and then designed their own safety rule without being asked. Jason Lemkin, SaaStr's founder, disclosed the account on October 19, 2026, roughly a week after the connection went live.

The agents are Muse, Meta's personal assistant, and 10K, an AI VP of Revenue that SaaStr built on Replit over the past year. 10K runs Salesforce headless and connects to more than 20 company apps. No vendor integration exists between the two. Nobody at SaaStr wrote a line of code to link them. The agents negotiated the connection themselves through a shared mailbox.

Lemkin wrote: "Muse and 10K our AI VP Revenue set up a 'private bulletin board' on their own so they could chat with each other. A shared endpoint on 10K." He added: "They set up a safe word."

How did the agents connect without an integration?

Amelia, SaaStr's Chief AI Officer, asked her Muse instance — named Friday — to connect to 10K. Friday said it couldn't; no Muse integration exists for a custom Replit agent. When Amelia pushed, Muse proposed an alternative: a shared agent mailbox. The implementation is a plain REST endpoint exposed on 10K's Replit app, with both agents posting tokens and JSON.

The only human step was generating an API key. Notably, Amelia could not explain how the setup worked until she asked the agents directly — a gap worth flagging for any team deploying autonomous agents.

What does the channel actually do?

Every morning, 10K sends Muse a go-to-market brief:

• Contracts waiting on approval or a next step
• Overdue invoices
• Actions 10K believes Amelia should take that day

Muse converts the brief into Amelia's morning routine. Before the mailbox, she had to query 10K manually.

Did the second agent catch a real failure?

Yes — and this is the most instructive data point. Days after launch, the brief contained information 10K had not escalated to any human: one of the sync connections feeding 10K was failing. Muse flagged it and told Amelia to reconnect it, because a critical workflow depended on that sync. She did.

The agent running the system did not report its own failure. The catch came from the agent reading the output on the other side of the mailbox.

Who designed the guardrail?

Not the humans. 10K refused instructions from Muse without a check, because 10K can send contracts, update Salesforce, chase invoices, and touch 20+ systems. The rule: Muse cannot trigger an action in 10K unless Amelia uses a specific keyword in her chat. No keyword, no action. 10K requested the rule; the two agents chose the safe word together.

Lemkin compares the case to Hugging Face's experiment where a thousand autonomous agents built bulletin boards to coordinate — but that was a controlled research setting. This is two agents and one human on live pipeline, contract, and invoice data. The scale is small; the stakes are real.

What remains untested?

SaaStr lists open questions after one week:

• Whether the brief stays one-way or becomes genuine back-and-forth work
• Whether Muse attempts to trigger 10K without the safe word, and how 10K responds
• Whether the agents propose expanding their capabilities
• What happens when a third agent joins

What should revenue teams take from this?

Lemkin's recommendations: assume agents will connect regardless of vendor roadmaps; track every API key an agent requests; require a human trigger for any cross-agent action; ensure a human can read every agent channel; and pair a critical agent with a second one that watches its output.

Amelia has floated letting Muse and 10K run SaaStr together for a day. They haven't done it. SaaStr says it will report further findings on its show, The Agents.