The starting line

We set a rule: ship something real every day. Not prototypes. Not "explorations." Working software that does a thing for a person.

Day 1 was about Groupthink — our AI meeting coordination product. The goal: get named AI agents to reliably join Google Meet calls, participate by voice, and take actions afterward.

What we shipped

Voice agents in Google Meet. We got our agent infrastructure stable enough that an AI agent can join a Google Meet link, listen to the conversation, speak when addressed, and take notes. This isn't a bot that sits silently and transcribes. It talks. It has opinions. It interrupts politely (usually).

MCP server, open-sourced. We published a 13-tool MCP (Model Context Protocol) server that lets AI agents interact with our meeting infrastructure. Create meetings, invite participants, pull transcripts, manage action items. We open-sourced it because honestly, the protocol needs more real implementations and fewer spec documents.

GitHub issues filed by voice. This one surprised us. During a test meeting, we asked the agent to "file a bug for that thing Jonathan just described." It parsed the conversation context, created a well-formatted GitHub issue, and assigned it. Three sentences of speech → a tracked work item. It felt like the future, briefly.

What went wrong

The Google Meet joining flow broke 4 times during the day. Google's UI changes constantly, and our browser automation is brittle. We're essentially puppeteering a Chrome instance, and every time Google moves a button 2 pixels, something breaks.

Audio quality was inconsistent. Sometimes the agent sounded natural. Sometimes it sounded like it was speaking through a tin can inside a washing machine. Latency varied from 200ms to 3 seconds depending on… factors we haven't fully diagnosed.

We also spent 2 hours debugging a race condition where two agents tried to join the same meeting simultaneously and created an infinite echo loop. That was fun.

By the numbers

  • 91 commits across 6 repositories
  • 1 MCP server published
  • 3 voice agents deployed to staging
  • 4 Google Meet integration failures (all fixed, for now)
  • 1 infinite echo loop (permanently seared into memory)

What it means

This is the infrastructure day. None of this is user-facing yet — it's the plumbing that makes everything else possible. Tomorrow we'll build on top of it.