Andrew Lee has a knack for spotting infrastructure gaps before they become obvious. A decade ago, he and his team built Firebase—the backend-as-a-service platform that became one of Google's quieter but more enduring developer bets. Now Lee's making another wager, and this one's considerably more audacious: that the software of the next ten years won't be written by developers at all, at least not in the traditional sense.
It'll be built by AI agents. And those agents will need something to run on.
That's the premise behind Tasklet, the San Francisco startup Lee co-founded with Jonny Dimond, one of Firebase's early engineers. On April 7, the company emerged from relative obscurity with a $20 million funding round and general availability of what it's calling an "operating system for AI agents." Union Square Ventures and Lightspeed Venture Partners co-led the round, which values the eight-person company at $175 million—a number that raised eyebrows even in a market that's grown accustomed to frothy AI valuations.
Y Combinator joined the round. So did Google's Jeff Dean and Stripe's Collison brothers, Patrick and John. When industry heavyweights pile in like that, it's worth asking what they're seeing.
The company claims $5 million in annual recurring revenue despite having barely existed in its current form at the start of the year. Growth, they say, has topped 1,200% since January 1, 2026. Union Square Ventures went a step further, publicly stating in an April 8, 2026 blog post that it had "rebuilt our entire operating system on the Tasklet stack" after becoming the company's first enterprise customer. That's either extraordinary confidence or extraordinary salesmanship—possibly both.
Beyond the Chatbot Prompt
Here's what Tasklet actually does, stripped of the hype: it runs AI agents that complete tasks rather than just responding to queries. The distinction matters more than it might seem.
Most AI tools today require constant human supervision. You ask a question, get an answer, then do something with that answer. Tasklet's pitch is that its agents operate more like junior employees—give them a task, and they'll work through it, even if it takes hours or days. They run continuously in the cloud, tapping into more than 3,000 pre-built integrations, any HTTP API, and something called MCP servers that help agents access external data.
Each agent operates inside its own Linux sandbox, complete with Python 3.12, ffmpeg, ImageMagick, and the ability to install whatever other tools a task requires. They can be triggered by calendar schedules, incoming emails, Slack messages, or webhooks. Then they keep going until the job's done.
When an integration doesn't exist—and there's always some obscure internal tool that doesn't have an API—Tasklet does something clever. It spins up a full cloud virtual machine with a graphical interface that persists between sessions, essentially giving the agent a computer to operate. In late April, the company shipped a built-in Chrome browser specifically for agent web automation, a faster and cheaper alternative to the full VM for tasks like navigating websites, filling forms, or downloading files.
The interface is deliberately spartan: a chat window backed by what the company describes as a "cloud-native agent runtime." But underneath sits multi-model support (GPT-5.5 was added on May 13), automated routing that sends transcription tasks to one model and video generation to another, and the ability to generate "Instant Apps"—small interactive web interfaces connected to live data streams.
It's slick. Whether it's reliable enough for real work is another question entirely.
The Infrastructure Play
Tasklet's architecture reveals something about where the AI infrastructure conversation has landed in 2026. The company is betting heavily on what developers call "durable execution"—the ability for long-running processes to survive failures, server restarts, and the general chaos of distributed systems.
This isn't a new concept. Frameworks like Temporal, Inngest, and Cloudflare Workflows have been tackling durable execution for years, treating reliability as fundamental rather than an afterthought. What Tasklet brings is a different delivery model—wrapping all that complexity in a managed service so engineers don't have to think about workflow definitions or state machines.
You describe what you want in plain language. The platform handles the execution, persistence, and recovery when things inevitably go wrong.
This is where the Firebase DNA shows most clearly. Lee and Dimond built their reputations on abstracting away infrastructure—making complicated backend services accessible through clean APIs. Tasklet applies the same philosophy to AI agents. State management, integration authentication, sandbox isolation, browser automation: all of it disappears behind the interface.
The company is pursuing CASA Tier 2 certification. SOC 2 and GDPR compliance are in progress, which signals enterprise ambitions that extend well beyond the early adopters currently running Slack bots and email triage workflows.
The Category Question
Tasklet occupies uncomfortable territory between two established markets, and that might be its greatest asset—or its biggest vulnerability.
On one side sit the workflow automation incumbents: Zapier, n8n, Make. These tools excel at deterministic, visual flow design. If this happens, do that. They're predictable, which is exactly what many businesses want. Tasklet argues this model collapses when tasks require judgment, when steps can't be mapped out in advance, when the workflow needs to adapt on the fly.
Its agents, in theory, handle ambiguity. They plan. They adjust. They don't require users to anticipate every possible branch in a decision tree.
On the other side are the developer-focused frameworks—LangGraph, various open-source agent runtimes—that offer fine-grained control at the cost of setup time and infrastructure headaches. Tasklet's bet is that most teams want results, not architecture projects. They want to describe a task and have it completed, full stop.
It remains to be seen whether this middle ground represents a defensible category or just a transitional moment before the market consolidates. The growth numbers are encouraging. But converting early traction into sustained enterprise adoption means proving that agents can be trusted with mission-critical work, not just the tedious stuff humans would rather avoid.
Moving Fast
The roadmap since general availability has been aggressive, perhaps more so than the founders expected. After launching the built-in browser in late April, Tasklet added GPT-5.5 support in May, then audio transcription and video generation capabilities that intelligently route tasks to specialized models—Gemini for transcription, Veo for video, GPT Image 2.0 for generation.
The pace suggests a team racing to define what an agent operating system should include before industry standards solidify and competitors catch up. Pricing, however, remains curiously opaque. The website mentions a refund policy for amounts under $250 and references live video support for higher-tier plans, though a complete public pricing grid is not available. Enterprise customers are directed to contact sales—a traditional B2B playbook that feels oddly at odds with the company's developer-first heritage.
Union Square Ventures' Nick Grossman described the investment as infrastructure for what he called a "post-workflow world." If he's right, Tasklet's challenge won't be building features. It'll be convincing businesses to trust agents with work that humans currently supervise, double-check, and ultimately take responsibility for when things go wrong.
The pedigree helps. Firebase wasn't an overnight success, but it became indispensable infrastructure for millions of developers. Lee and Dimond know how to build tools that stick around.
Now they just need to prove that agents are ready to run the shop. Or at least parts of it.
