How Shortwave Wants To Reinvent Email With AI
For better or worse, it won’t be too long before virtually every productivity application you use will include some form of agentic AI. To some degree, email apps are obvious candidates here, given how central they are in most users’ day-to-day work. For the second episode of The New Stack Agents, I spoke with Andrew Lee, co-founder and CEO of Shortwave, a Gmail-centric email client (with Outlook support coming soon) that initially took a more traditional approach to managing inboxes before pivoting to put AI at the forefront of the user experience.
From Firebase to Shortwave
Before founding Shortwave, Lee was one of the co-founders of Firebase, the backend-as-a-service platform Google acquired in 2014. It was at Google, Lee told me, where he realized that he wanted to build a better Gmail.
“We started in 2020 and the initial plan was just to build a better Gmail,” he said. “We had a lot of fairly traditional ideas about how to build a better product experience, like we wanted to change the way that threads were displayed, we wanted to give you better search, we wanted to give you some cool collaboration and sharing features.”
Lee argued that, for structural reasons, it was very hard for Google to innovate on top of Gmail, given how many stakeholders there were and how hard it would be for its two and a half billion users to adapt to those changes.
Maybe one of the most interesting things Google did at the time was launching Inbox in 2014 to overcome exactly this problem. It was a new, completely rethought email client that used Gmail as its data source. This allowed the team to experiment with new ideas without touching the traditional Gmail experience. Despite its large fan base, Inbox fell victim to Google’s cost-cutting efforts in 2018 and Shortwave, to some degree, is resurrecting some of that original ethos behind Inbox.
“There were several people on the Firebase team that were like refugees from the [Inbox] team before they actually shut it down, but kind of saw the writing of the wall there,” Lee said. “And then the folks who were left on the Gmail team had a very clear signal of, like, don’t try to innovate too much, right? If you innovate too much, your project might get killed. And so, you know, if you look the last six years since that product got killed, Gmail hasn’t done a lot. They’ve kind of shoved the Gemini sidebar in there, but there’s been no radical efforts to rethink the product.”
Putting LLMs Into Your Inbox
By the autumn of 2022 and with the launch of ChatGPT, Lee said, it became clear to him that large language models (LLMs) were going to be a thing. But at the time, they were still slow and expensive, so what the team did was add a few LLM-powered features like summarization, translation, and listing action items.
“All of [these features] were like, you click a button or select a menu option, and then a modal pops up and you wait and a thing gets spit out. And honestly, none of them were terribly useful, except maybe the summary feature. But people talked a lot about it, right? There was a lot of hype. We got a bunch of press from it. We got a bunch of signups. A bunch of people paid us for it. Definitely enough excitement that we decided to double down,” Lee explained.
Fast-forward to today and the models have gotten better, faster and cheaper to run, but maybe even more importantly, they’ve gotten better at using tools to get their work done. The Shortwave team actually started implementing a form of tool calling early on, using their own DIY approach. Then, when the model providers like OpenAI started implementing tool calling into their APIs themselves, the early versions weren’t all that good (“We tried it, and what we discovered at the time was that it felt like anytime we had it run a tool, it actually got dumber in general”).
It was the launch of Anthropic’s Claude Sonnet 3.5 last year that changed the game for Shortwave.
“It seems like Anthropic has been doing some very focused reinforcement learning, specifically on agentic tool calling use cases, and last summer, at this time, the OpenAI models, they could effectively call like one tool, but you couldn’t really get them to call a bunch of tools, or a sequence of tools. And the Anthropic models, starting in December, were able to call a tool, look up information, take that information to figure out the next tool, call another tool, and do that in sort of an endless sequence. And the result of that was a radically smarter agent.”
For tool calling use cases, Lee argues, Anthropic still remains far ahead of the competition, especially now with the rollout of the latest 4.0 generation of models.
The one thing that still needs work, though, he noted, is that once you add too many tools (and that mostly means Model Context Protocol servers at this point) and integrate too many third-party services (which themselves may be calling their own tools), things start to become problematic and the reasoning starts to break down. “I think we really need models that can handle a thousand tools without breaking a sweat,” Lee said.
Looking Ahead: Dynamic UIs and Proactive Apps
Looking ahead, one of the areas the Shortwave team is working on is having the AI proactively take actions on the users’ behalf. That remains a costly proposition, though, given that every incoming email would likely result in lots of token usage.
Another area he is excited about is being able to generate the right user interface for the right task on the fly. Coding agents, after all, can generate user interfaces on demand, but what if that didn’t just happen during the coding stage but live, within the app, so that when you are reviewing candidates for a job, the inbox would look different from when you’re working on handling customer support.
