Key Takeaways
- Autonomous software agents are turning into the primary consumer class that evaluates, tests, and purchases developer infrastructure.
- Existing developer platforms still force AI agents to impersonate humans with personal email addresses and manual account verification to obtain API keys.
- WorkOS built an internal authorization engine called Airlock to handle dynamic permissions and registration specifically for AI agents.
- Frontier engineering teams already deploy agents like Atlas, Devin, and GroqBot directly inside Slack and codebases as autonomous team members.
Software Agents Are Becoming the Primary Buyer
For two decades, developer-first companies won by bottoms-up adoption. You wrote clear documentation, built a self-serve API, gave away a free tier, and waited for individual engineers to swipe a credit card. That playbook is shifting beneath software founders.
Michael Grinich, founder and CEO of WorkOS, argues that software agents are taking over technical stack decisions: “So like one thing that you see with developers, especially developer focused companies, is the kind of acceptance that agents will be the ones choosing our technical stack, and kind of becoming the like consumer class to sell to.” (play from 30:04)
When an AI agent writes code, resolves GitHub issues, or spins up cloud services, it does not browse marketing homepages or read blog posts. It inspects documentation, checks endpoint latency, evaluates API schemas, and selects whichever service lets it accomplish its objective with the least friction. If your product requires human interaction to onboard, agents will simply choose your competitor.
The Machine Identity Bottleneck
Every modern onboarding flow is built on the assumption that a human sits behind the screen. You ask for a work email, verify a phone number, send a magic link, and run a CAPTCHA. When an agent attempts that flow, the system grinds to a halt.
Grinich points out the absurdity of current workarounds: “Like if you build an, you know, an infrastructure service and you want to allow an agent just to go register, without this, the agent has to have an email address or somehow verify its identity somehow, or it has to impersonate a human.” (play from 30:26)
To bridge that gap, WorkOS developed an internal mechanism called Airlock. “I imagine some things will come out of that and turn into commercial products, like another one I can talk about in a in a in a bit is Airlock, which is our agentic permissioning system. It's like a new thing for dynamic permissions for AI agents,” (play from 13:53) Grinich said. Instead of granting an agent static API keys with broad administrative rights, agentic systems need temporary, dynamically scoped permissions tied to verified machine identities.
Frontier Teams Treat Agents as Teammates
Most software companies still treat AI as a chat widget or a text completion box. Frontier teams run in the opposite direction, treating agents as distinct operators with their own accounts, communications, and permissions.
As the The Changelog hosts noted during the discussion: “The the realness of that frontier, I feel like that frontier is is accepting, because you kind of began with this this identity factor, right?” (play from 36:45)
Grinich sees this shift happening among builders running tools like Devin, Claude in Slack, or GroqBot: “I think right now the idea of agents being a big part of the workforce is like this. If you spend time on X, you'll see this with stuff like WorkOS Atlas, certainly the coding agents, you know, like people with Devin or Claude Tag or things inside of Slack, and now with GroqBot, the the GroqBot product. But I think that's still not common knowledge or accepted yet.” (play from 31:08)
WorkOS built Atlas as an internal teammate rather than a side tool. Once you grant an agent its own workspace identity, the engineering challenge shifts from model output quality to machine authorization. Builders who solve agent onboarding and identity today will capture the autonomous buyers of tomorrow.
What to Do With This
Run an audit on your product's signup flow tomorrow using an autonomous script. Attempt to register an account, provision API credentials, and query your core endpoint without clicking an email confirmation link or completing a CAPTCHA. If your workflow blocks the script, design an authenticated machine registration path with scoped API tokens so software agents can onboard without a human handler.