I want to tell you about Eddy (Headway's AI agent/stack): Last month I merged 60 PRs. Not because my calendar suddenly cleared, but because the way I work right now looks nothing like it did a few months ago. AI as a "copilot" is already outdated. When it’s actually wired into your systems (and built with real safety/constraints), it stops being a tool you use and starts being something you work with. You try more things. You ship more. You stop overthinking whether something is “worth the time” and just do it. A small team at Headway decided not to wait around for that to exist. So we built our AI stack. What’s been most interesting isn’t the speed, it’s the behavior change. Engineers aren’t just moving faster; they’re attempting things they would have previously filtered out as “too heavy.” Non-engineers are building workflows that used to require dedicated teams. The line between idea and execution is getting very thin, with that people attempt bolder things. I knew we'd built something special when our internal beta grew from 10 people to 600 in a month. Hard to really explain in just one post, so I wrote more on our blog about what we built, the constraints (especially in health care), and what’s coming next. Read it here: https://lnkd.in/e6iVTtvZ If you’re excited by building systems like this, we’re hiring. #AI #Engineering #Healthcare cc Headway Adam R. Kuhn Francisco Alvarez Matthew Tse Mike Perrone Amish Gupta Hagen Green
The data context registry is the most underrated piece of this entire stack. Everyone focuses on the agent and the MCP connectors, but the reason non-engineers are pulling their own reports isn't the AI - it's that someone finally codified the business rules that used to live in Slack threads and people's heads. That's the real unlock. Without it, the agent just confidently returns wrong numbers. The PHI classifier on the sharing layer is smart too. Most teams bolt compliance on after the fact; building it as a native layer means you're not choosing between velocity and safety every time you ship a feature. One question - with Snowflake as the warehouse and MCP as the connector layer, how is the context registry structured? Is it metadata tables inside Snowflake itself, or a separate service that Eddy queries before hitting the warehouse? Curious because the pattern of 'teach the agent the business rules before it touches the data' is something more teams should be stealing.
sarah Lange
3mo
This is such an exciting development in healthcare. As someone who has spent years in financial operations supporting large organizations, I know firsthand how much a well-integrated system can change the way a team works. Watching the line between idea and execution get thinner is exactly what innovation looks like. The fact that your internal beta grew from 10 to 600 users in a month says everything. Mental health access is one of the most important problems we can solve right now and it is exciting to see AI being used responsibly to help make that happen. Congrats to the Headway team on building something truly meaningful.
Kodi Chouinard
3mo
"Exciting to see how Headway is leveraging AI with Eddy to streamline the provider side. The next frontier is definitely the Engagement Gap—how we use similar AI logic to capture and stabilize higher-risk demographics (like 18-30 males) before they even reach the intake stage.
Mohi A.
3mo
Hey, I just read through the post—really impressive stuff. It’s cool to see a team actually move past the 'copilot' phase and build a true agentic stack that handles the messy reality of healthcare compliance and data context. Seems like a great product that’s actually changing the way people work. Awesome work on this!
Salwa Laaraichi
3mo
THIS! People are not just moving faster, they are attempting things they would have previously ruled out as too heavy. That is where AI starts becoming genuinely valuable. From an operator perspective, I’ve been exploring that on Hugging Face too: AI is most useful when it helps more people explore, test, and solve problems with more rigor, not just generate faster. Really impressive to see Headway building that with real constraints in mind. Feel free to check out my profile as well.https://huggingface.co/Salwa5
Henric Zhang
3mo
Why is it called Eddy? And not ed or edd
Luke Kujacznski
3mo
You should really connect with Shirali N.
Adrián V. take a look, I think you'll find this super helpful!
Blake Yoder
3mo
amazing!
This is a masterclass in why governance has to be an embedded control plane, not just a policy document. The most insightful part is the drop in cost of trying things, which only happens when teams own the workflow rather than rent it from a vendor. At Keywell, we're seeing the same pattern in health plans: when you wire agents directly into the environment, AI shifts from a point tool to an operating layer. Question: As Eddy becomes "more" agentic, how are you managing the lifecycle of the PHI classifier, and what controls ensure its sharing layer remains auditable as connections scale?