Key Takeaways
- WorkOS built Atlas (AI Teammate Learning and Automation System) inside Slack so engineers avoid jumping between tabs to execute automated tasks.
- Coworker personas do not make underlying models smarter; Grinich explains that teammate interfaces exist purely as ergonomic accommodations for humans.
- Instead of relying on one general assistant, WorkOS runs a few hundred narrow sub-agents inside Atlas for specific sub-projects and operational jobs.
- WorkOS Labs built Atlas as an internal experiment before deciding to release it publicly to solve workflow drag for other teams.
Coworker Personas Are Human Crutches
Most AI tools force workers into separate browser windows. You open a chat tab, type a prompt, copy the response, and paste it back into your group channel. That friction kills momentum. At WorkOS, the remote team worked primarily in Slack, so they built Atlas directly into channel threads.
Grinich is clear about why Atlas mimics a human colleague. Giving an agent a name and an avatar does not improve its underlying reasoning. As Grinich noted: “The the reason to embody it sort of in something like Atlas is not because it makes the AI better. It's because it makes it easier for us to connect with.” (play from 38:51)
Human brains evolved to work with other people through dialogue and shared documents. When software mimics those patterns, team members immediately understand how to assign work, inspect answers, and correct errors. Placing the model in a public thread lets the whole team see the work, check the results, and guide the system openly.
Narrow Sub-Agents Beat Monolithic Bots
Early adopters often ask one chatbot to handle all company knowledge. That setup falls apart because general prompts miss context and fail on specific technical tasks. WorkOS solved this problem by dividing duties across small components.
“Atlas is actually an acronym. It stands for, AI Teammate Learning and Automation System,” (play from 12:24) Grinich explained. The core system provides the shared interface, but the actual work goes to targeted sub-agents. “And you can build sub-agents inside of Atlas for specific tasks, or specific sub-projects. So we have a few hundred of those that do you know, more narrowly scoped things.” (play from 13:00)
Instead of forcing one bot to understand every repository and database, you scope each job. One sub-agent checks deployment statuses. Another pulls customer logs for support investigations. When someone tags Atlas in a thread, it runs background automations without forcing engineers into third-party dashboards. Grinich said: “And the idea is, you interact with it it's it's like a coworker. So you can tag it in threads, you can tag it to do anything with you. And it learns.” (play from 12:37)
Build for Internal Pain First
Atlas started inside WorkOS Labs, an experimental group inside the business. Grinich explained their focus: “We have a team inside of WorkOS called WorkOS Labs, which is kind of building these like new frontier experiences where we're not quite sure if we'll commercialize them or not.” (play from 13:38)
WorkOS built Atlas to remove their own operational bottlenecks. Only after operating hundreds of sub-agents internally did they release the tool publicly. If you want agents that your team will actually adopt, ignore generic templates. Find where your developers copy data between tools, place the bot directly into that communication stream, and shrink its responsibilities until it runs reliably.
What to Do With This
Audit your team's primary Slack channels from yesterday and find the three most common status questions engineers asked each other about deployments, pull requests, or bug reports. Pick one of those queries and build a single-purpose Slack webhook that answers it when tagged in the thread, instead of asking your team to log into an external dashboard.