[All writing]

Tooling That Shows Its Work

Jeff Liu//2 min read

For the past few months I've had 91 Trigger.dev tasks running content and social pipelines, including AI sentiment analysis on what's actually being said across publishers and social platforms.

My read is that most AI products stay black boxes on purpose, the same way privacy policy emails are built to go unread.

Compliance work in adtech and fintech trained a habit into me. When disclosure language is part of the job, you stop treating it as noise. That eye doesn't turn off once you change industries.

It's why building on Trigger.dev stood out. The observability is real, custom metadata and pipeline telemetry, queryable on every run, whether it's an AEO pipeline, image generation, or sentiment analysis. Claude sits inside that same visible pipeline, which is most of why latency dropped on the tasks I ran through it.

Their MCP connector is what actually taught me the infrastructure. Claude is only as good as the context you give it, and their documentation preloaded through that connector gave mine the context to build systems I had no idea how to build on my own. I broke a lot of it along the way. The dashboard let me look at that trial and error under a microscope, watch what the data actually did, run by run. That's where the confusion became better AI observability, and it's what let me go back and finally understand abstractions from earlier in my career that never fully clicked at the time.

They also built their MCP server early, and shipped it open source. Worth crediting on its own.

Refactoring my stack on tooling like that beats maintaining a black box. I stop guessing what a pipeline actually did and start reading it.

If you're running AI systems at volume and want the pipeline to show its work instead of asking you to trust it, worth a look.

They just shipped chat agent too, durable AI chat that runs as an actual stateful machine instead of a fragile API route, no timeouts, keeps streaming through crashes and refreshes. If you're building a chat-based agent and don't want to own that infrastructure yourself, that's worth a look on its own.