Trigger.dev Pivots to AI Agents, Hits PMF with 90% Usage

Product Evolution from Async Jobs to Agent Infrastructure

Trigger.dev's founders Matt and Eric describe three major iterations before finding product-market fit (PMF). Version 1 launched pre-YC in February 2023 as "Zapier for developers," an async background jobs framework targeting back-office automation like GitHub tasks or marketing workflows via Zapier/n8n replacements. It gained traction with a strong Hacker News launch, praised for its design-forward approach—code snippets first on the landing page, intuitive SDKs emphasizing developer experience (DX) over just visuals. Co-founders Dan and James drove UX, ensuring SDK functions minimized failure points.

Version 2 shifted to embedding async tasks directly into customer products, moving beyond internal tools to user-facing value like document processing or video encoding in the app's hot path. Adoption was decent but lacked PMF; serverless trends made long-running tasks harder, yet the product still required messy customer-side code execution. A customer poll revealed 60% assumed Trigger handled execution, highlighting the need for full offloading.

Version 3, released June 2024, delivered: an SDK + platform + infrastructure executing TypeScript-first code on Trigger's machines. This unlocked reliability for retries, idempotency, and long runs—perfect for AI. Revenue surged 30%+ monthly post-beta pricing, confirming PMF immediately. Over 90% of current usage is agent workflows, accidentally positioned by two years of async infra building just as AI agents exploded.

Real-World Agent Use Cases Driving Growth

Customers pull Trigger toward agentic patterns with human/LLM-in-the-loop. Icon.com replaces video ad agencies: users upload product assets; Trigger processes/classifies them in real-time (visible progress), generates hundreds of ads via AI (e.g., actors), accepts feedback, and posts to TikTok/Instagram. This splits into context gathering (asset processing) and action (prompt-based generation), pausing for input.

Magic School (edtech) powers teacher tools for lesson planning, Q&A, grading, and student catch-up via fully Trigger-executed agents. Scrappy Bar's coding agent hooks GitHub, pulls code, iterates with LLMs (e.g., evaluating changes), runs tools like Puppeteer/FFmpeg/Chromium for browser automation, and commits—all on Trigger machines with shell/Python access.

These showcase two agent phases: context building (stateful, long-running) and execution (dynamic, feedback-driven). Trigger pauses compute on feedback, resuming seamlessly—human or agent-provided. Developers get full machine power via TypeScript SDK, no infra worries.

Open Source Strategy and LLM as Users

Apache 2.0 licensed, nearly everything open-source except Kubernetes-scale management (Helm chart available, but hard to scale). This aids monetization via cloud compute billing, as devs avoid infra. Open-source footprint boosts LLM recommendations (e.g., Claude suggesting Trigger), with repos enabling self-debugging via AI: customers paste Claude-found bugs, which auto-PR fixes.

Post-Claude 3.5 Sonnet ("Opus 4.5"), "vibe coders" (non-traditional devs) blurred with pros—better models + Trigger's LLM-friendly docs/MCP servers/skills reduced support bifurcation. Support volume grows slower than metrics; open-source lets LLMs "read tests" for truth. Marketing eyes agent onboarding: Claude could soon self-spin accounts end-to-end.

"There's kind of two users now: the human user who wants to build something but also like the LLM is a user of Trigger," Matt notes. Open-source maximizes this, dwarfing closed-source visibility.

Future Vision: Programmatic Checkpoint and Restore

Founders predict computing's future as "programmatic checkpoint and restore"—freezing/resuming full machine state (CPU, memory, filesystem) on demand, like an OS scheduler over containers. Traditional rehydration is brittle; Trigger democratizes low-level snapshot tech for agent loops, pausing mid-workflow until events (feedback, data) trigger resume.

This powers reliable, stateful agents without dev overhead, positioning Trigger as agent infra. Luck from async focus met AI timing perfectly.

Hiring, Quality, and Founder Advice

Post-Opus 4.5, hiring emphasized agent workflow quality: shipping robust code despite vibe coders. Advice for new founders: obsess over first 5 seconds (code-first landing), design DX deeply (hard SDK debates), pivot boldly (pre/post-YC), value design in devtools. YC batch honed landing page/code snippet focus; raised post-batch, Series A ($16M, Standard Capital) followed PMF.

"We spend probably the hardest conversations we have around the product are like how to design this specific SDK function and make it so it's very very hard for developers to like fail when they use it."

Key Takeaways

• Launch code-first: Prioritize SDK simplicity and landing page code snippets to hook devs in 5 seconds.
• Pivot to execution: Offload async code runs to your infra—60% users expect it.
• Split agents: Build context (stateful processing) separately from actions (dynamic generation) with pause/resume.
• Open-source for LLMs: Maximize repo visibility so AIs recommend and debug your tool.
• Checkpoint-resume future: Programmatically freeze/resume machine state for reliable agents.
• Design DX holistically: UX + API design prevents failures; British co-founders nailed it.
• Track PMF signals: 30%+ MoM revenue post-version 3 confirmed agent fit.
• Blur vibe/pro coders: LLM advances + docs make tools accessible without support explosion.
• Use cases lead: Icon/Magic/Scrappy show video gen/edtech/coding agents.

Notable Quotes:

• Matt: "Over 90% of usage now comes from agent workflows."
• Eric/Matt: "We think the future of computing is basically programmatically being able to do this type of snapshot and restore mechanism."
• Matt: "Code first approach... I always hate going to someone's developer website and having to scroll scroll scroll and there's no code."
• Pete/Matt: "It almost seems like this thing that you had been building for almost two years... was just perfectly designed for a use case that didn't actually exist at the time."
• Matt: "The LLM is a user of Trigger... being open source is a massive advantage there because we have a much bigger footprint on the internet."