Finance, 11 November 2026
Tasklet: Five Engineers, Claude Code, and $2.5M ARR in Five Months
Tasklet — a cloud AI-agent platform for knowledge work, built by a five-engineer team under founders Andrew Lee (who previously started Firebase and the AI email client Shortwave) and Jonny Dimond (a former Google staff engineer). Tasklet grew directly out of an automation feature inside Shortwave, was rewritten from scratch in mid-2025, and launched as its own product in October 2025.
Problem
Older workflow-automation tools (think Zapier-style “if this, then that” builders) need a human to map out every branch and edge case in advance, and they can’t act on anything that wasn’t anticipated. Shortwave’s own users, meanwhile, kept asking for AI that would run on its own — triggered by an email or a calendar event — instead of waiting for someone to open a chat window and ask. The blocker had been the models themselves: strong at answering a single question, but unreliable once a task stretched across many steps and tool calls.
Result
$2.5M
- $2.5M in annual recurring revenue within five months of its October 2025 launch, growing roughly 160% month over month.
- 450,000 agent actions executed per day across customer accounts, with about 2,000 new agents created daily.
- By its January 2026 run-rate, growth had compounded further — Y Combinator’s own profile puts it at roughly 1,200% growth since the start of the year, reaching about $5M ARR and a $20M Series A (led by Union Square Ventures and Lightspeed, at a $175M valuation).
- Customer-built agents include email triage, Stripe/Ramp revenue reconciliation, competitor monitoring, and — built by one customer in under a week — a full multi-agent back-office system for a venture fund.
Solution
Tasklet lets the underlying model reason through a workflow step by step instead of following a fixed script. Every call goes to Claude — Sonnet 4.6 for routine work, Opus 4.6 with extended thinking for anything complex — and each agent is really just a long-running Claude conversation with three things attached: over 3,000 pre-built integrations (plus new connectors it can build on the fly from a service’s API documentation), triggers such as schedules, incoming emails, webhooks and calendar events, and a cloud browser for the services that have no API at all. The five engineers who built all of this wrote it using Claude Code for their own internal development, start to finish.
Could a small team do this?
Partly, and the honest caveat matters here: Tasklet is itself proof that five engineers can build a serious AI-agent product on Claude Code — but what they built is a general-purpose platform, not a single tool, and that’s the harder part to copy solo. The 3,000+ prebuilt integrations, the cloud-browser fallback for API-less services, and the safety and reliability engineering needed to let agents act autonomously without a human watching every step each took real engineering months, backed by $20M in funding. What a solo operator can realistically copy is the narrower pattern underneath it: build one agent, with Claude Code, wired to a handful of APIs and a trigger (a schedule, a webhook, an inbox), aimed at one specific recurring task for one type of customer — the “mini-Tasklet” rather than the platform. Email triage for a specific industry, or invoice reconciliation against one accounting tool, is a realistic few-weeks build; a 3,000-integration platform that creates 2,000 new agents a day is not.