David Crawshaw is an Australian programmer and infrastructure founder whose career tracks three eras of developer tooling: Google's cluster-scale systems and the Go language, the mesh-VPN startup Tailscale, and — since late 2025 — exe.dev, a cloud built for a world where software is increasingly written by AI agents.

## From North Queensland to Google

Crawshaw grew up in a small town in North Queensland, Australia, and learned to program about the time he learned to read, on the computers of his parents' medical practice — MS-DOS machines wired to a Novell NetWare file server, where his father moonlighted writing the practice's medical-record software in Clipper. He has described it on the Kubelist and High Leverage podcasts as the ideal small business in which to learn programming, and an accidental education in product design: watching receptionists check patients in, and asking how to make it take less than a minute.

He studied mathematics and history at university, moved to the Bay Area for family reasons, and landed at Google — in his telling, his first real US job — around 2009. He wanted clusters: work that could not be done on one machine. He got them. His first project moved petabytes between data centers on a logs team supporting search quality, in the Borg era before Kubernetes was usable inside Google. He stayed for the better part of a decade and left as a staff software engineer; his later Google work included implementing TCP/IP networking for the Fuchsia operating system.

## The Go years

His most durable public work at Google was on Go itself. In 2014 he authored the change adding GOOS=android to the Go runtime, and Go 1.4 — released that December — shipped official Android support built on the golang.org/x/mobile repository. He led the iOS port too, turning Go into a mobile platform through the gomobile tooling and the gobind binding generator. In 2016 he implemented the plugin build mode for the go command, and at GopherCon 2017 in Denver he presented "Go Build Modes," a survey of every way a Go program can be built to interop with other languages and itself. In 2017 he also released Neugram, an experimental scripting language and shell reusing Go's syntax and type system.

On March 27, 2018 he posted "Leaving Google": last day, mid-Fuchsia, off to do childcare and start a software business.

## Tailscale: a weekend hack that became a company

The business became Tailscale, co-founded in early 2019 with Avery Pennarun and David Carney — three former Google engineers, distributed across New York, Montreal, and Toronto from day one. The founding story he tells is deliberately unheroic: they knew the shape of the problems they wanted — computing made too difficult for no good reason — found a small Canadian bank that needed two-factor access to old .NET systems, and hacked up a program that wrote WireGuard config files in a weekend. It ran on a machine in the closet of his Manhattan apartment.

The product's architecture came from constraint. Working out of the New York Public Library — which blocked UDP — he collapsed their VPN-concentrator design into one multi-tenant control plane in the cloud with devices connecting peer-to-peer on the data plane, plus a TCP relay fallback. "Mesh networks never work," he remembers Pennarun replying, "but that sounds fine for now." It worked. Tailscale became the developer-beloved way to make a private network; in April 2025 it raised a $160M Series C at a reported $1.45B valuation. By then Crawshaw was leaving — he describes hiring himself out of a job and stepping back, fully, at the Series C.

## Sketch, agents, and exe.dev

What pulled him out was LLMs. Around late 2024 he and Josh Bleecher-Snyder started a company to improve developer tools with them, and ran through the era's product space at speed: a Go-specialized code-completion system, then Sketch — a coding agent that worked in a Docker container and returned results as git branches. Claude Code shipped a simpler agent and ate the market; users bounced off Sketch. But every company they tried it with had test suites that broke the container sandbox, so they kept rebuilding the substrate — gVisor, then real VMs — until the realization landed: the infrastructure was the product.

exe.dev launched at the end of December 2025 selling pools of CPU, memory, and local-NVMe disk on which customers run as many VMs as they like, on machines the company racks itself. On April 22, 2026 he announced a Series A and total funding of $35M, with Amplify, CRV, and HeavyBit named as investors; Amplify's own post frames it as backing "the perfect little computer." A web-based agent, Shelley, ships in the default image so users can build apps from a phone.

## How he thinks

His essays are the record's spine. Three commitments recur. First, empiricism: he spent a year deliberately learning what LLMs are good for — autocomplete, search, chat-driven drafts — and kept publishing revisions as the frontier moved ("Eight more months of agents" reports frontier models writing nine-tenths of his code, all of it carefully read). Second, a definition of agents as feedback loops: an agent is a for loop containing an LLM call with tools, and environment — compiler, tests, shell, browser — is what turns a demo into a programmer. Third, the inversion of agent infrastructure: agents are trained on how developers work, so "the best software for an agent is whatever is best for a programmer."

His most recent argument is institutional: review-then-commit code review cannot survive agents, which double review load while destroying the reviewer's ability to infer effort — a principal-agent problem he says is "killing OSS" through slop PRs. His practiced answer is small high-trust teams where the human driving the agent owns deployment; he reports it working with nine people at exe.dev, and is candid that it does not obviously scale to low-trust companies.

Underneath the timeliness is a longer continuity. The person who ported a language onto phones, replaced a VPN concentrator with a mesh, and now sells VMs-as-processes keeps making the same move: strip the abstraction until the actual computer is visible, then make it pleasant to use. Asked why he would found a second company after Tailscale, his answer is not a mission statement — it is "I like computers."

## What the record does not settle

The public record is thick on work and thin on the person: no Wikipedia article or Wikidata item exists, and dates like his Google start are approximate in his own telling. The Tailscale departure and the Sketch-to-exe.dev boundary are reconstructed almost entirely from his 2026 podcast appearances — self-narrated, plausible, and uncorroborated in the catalog. Even exe.dev's funding is stated two ways across his post and his investor's. This index preserves those seams.

*This index was compiled from public sources and does not imply the subject's endorsement. Citations live in the packet's source catalog.*