Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
From Zapier for Devs to Powering 90% AI Agents
TL;DR
Trigger.dev found PMF on its third version, not its first — after starting as “Zapier for developers” in YC W23, the company only really took off when it moved from an async SDK to fully executing long-running jobs on its own infrastructure in June 2024.
AI agents turned a decent infra tool into a breakout product — Matt and Eric say over 90% of Trigger.dev usage today is agent workflows, because agents need exactly what Trigger built: long-running execution, pause/resume, retries, and human-in-the-loop feedback.
The key technical shift was checkpointing compute, not just queueing jobs — Trigger’s big insight is that future agent systems need to freeze and restore full machine state (CPU, memory, filesystem) instead of forcing developers to constantly rehydrate state by hand.
Open source became a distribution advantage for LLMs, not just humans — because Trigger is Apache 2 and its repo, tests, and docs are public, Claude and other coding agents can understand the product better, surface bugs, and increasingly recommend it to users.
The company’s hiring plan changed materially after better coding models arrived — after raising a Series A in November, they expected to hire many more engineers, but improvements from tools like Claude Code and Opus 4.5 made each engineer feel 5x–10x more productive.
They now evaluate engineers on AI fluency, not pure unaided coding — Trigger’s interview process includes a paid trial day where candidates are expected to use AI tools well, because the founders see the modern job as leveraging agents effectively, not proving you can work without them.
The Breakdown
From YC’s “Zapier for Developers” to something much bigger
Matt and Eric open by describing Trigger.dev today as a way to add AI agents to your product via an SDK, with Trigger handling reliable execution. But that’s a long way from what they pitched in YC W23, when the company was framed as “Zapier for developers” and launched early in the batch as an async background jobs framework. The host remembers their Hacker News launch and, just as vividly, how obsessively they tuned the landing page and code snippet to win the first five seconds of a developer’s attention.
Design mattered — and for them, code was part of the design
The founders make a sharp distinction between visual polish and actual developer experience: design means making it “very, very hard for developers to fail.” They obsessed over SDK function signatures and even the hero code snippet because developers want to see the code first, not scroll through marketing copy. Their framing is telling: they designed the code they wanted users to write before worrying about how to make the backend work.
Version one and two worked, but didn’t quite click
The early product mostly got used for “back office” automation — GitHub workflows, sales, marketing, internal tools — basically the kind of operational plumbing Zapier already serves. That led them to version two, focused more on embedding async tasks directly into customer-facing products: processing documents, encoding videos, doing useful work in the product’s hot path. Adoption was fine, but not real product-market fit; as they put it, the market existed, but the product didn’t match the problem closely enough.
Version three: running the code themselves changed everything
The breakthrough came in June 2024 when Trigger stopped asking developers to execute jobs on their own infrastructure and started running the code itself. They even found in a customer poll that 60% of users already assumed Trigger was handling execution, which was a pretty loud hint. Once they made the switch, growth snapped into place — they say revenue grew at least 30% month-over-month for a long stretch, and PMF showed up almost immediately.
AI agents were the perfect use case they accidentally prepared for
What they had built for background jobs turned out to be exactly what AI agents needed: long-running workflows, lots of waiting, and reliable pause/resume behavior. Eric describes two parts of successful agents — gathering context and then acting on it — and gives Icon.com as the example: users upload brand assets, Trigger processes and classifies them, then generates hundreds of video ads for TikTok and Instagram. The sticky part is human-in-the-loop: the workflow can pause, wait for feedback from a person or another agent, then resume without losing its place.
Why 90% of usage is now agent workflows
Today, more than 90% of Trigger’s usage is agentic, with customers like Magic School for education workflows and Scrappybar for code-generation agents that interact with GitHub. Under the hood, developers get the power of a full machine image — shell access, Python, ffmpeg, Puppeteer, Chromium — even though the interface is TypeScript-first. Their bigger bet is that checkpoint-and-restore becomes a primitive of modern computing, almost like an OS scheduler for cloud containers, but abstracted away so developers never have to think about it.
The customer is now partly human, partly LLM
One of the most interesting moments is when they say there are now “two users”: the human builder and the LLM using Trigger on the human’s behalf. Six months ago they could clearly distinguish “vibe coders” from traditional developers through support requests, but better models, improved docs, MCP support, and more agent-friendly tooling have blurred that line. Because the product is open source, customers and agents can inspect the repo and tests directly; sometimes users even report bugs with Claude-generated writeups, and Matt can ask Claude to patch and open the PR.
Smaller teams, different hiring, same YC advice
The rise of coding agents changed their company plan fast: after their November Series A, they expected to hire many more engineers, but now say each engineer may be 5x–10x more productive. They’re still hiring, just less aggressively on engineering and more on roles like devrel and content, especially to teach people how to build better agents. Their hiring test is a paid trial day where candidates are encouraged to use AI however they normally would — the opposite of LeetCode — and their closing advice to new founders is classic YC: ship early, stay close to customers, and trust the signal if you’ve personally felt the pain deeply enough to keep going through multiple imperfect versions.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
