CTO @trymultiplayer: the debugging agent for developers.
Brooklyn, NY
Joined November 2007
- Tweets1.8K
- Following1.1K
- Followers279
- Likes2.2K
Pinned Tweet
Replace PR slop with a handful of high-quality PRs.
Coding agents prompted by incomplete observability data produce noisy, low-quality PRs that create more review work than they save.
Multiplayer provides coding agents unsampled, full-stack session data. 👇
Most companies are solving the AI quality problem at the wrong layer.
They're looking for better models or coding agents.
The problem is the data underneath them.
I know leaders who live in constant fear that a business-killing bug is lingering in some unread, AI-generated corner of the codebase.
Grounding non-deterministic AI output in deterministic input is just one layer of the solution.
I think it’s emotionally exhausting to manage an infinitely hungry AI workforce while knowing there is no longer direct human ownership over every line of code.
Tokenmaxxing was always going to end. Goodhart’s Law doesn’t fail.
What replaces it is the more interesting question.
Had a demo call earlier today where the CLI wizard set the live demo up in under a minute.
The user's reaction said it all.
Very proud of the team for building something that actually works that smoothly. That kind of experience doesn't happen by accident.
Show me a team's review queue and I'll tell you how they're using AI.
Reasonable revert rate, reviewers who feel like they're adding value: coding agents being used well.
I posted a screenshot of an AI-generated PR with 50k lines of changes. The comments exploded with opinions on PR best practices. Almost nobody stopped to ask whether the review process itself still makes sense.
It's like walking past a burning building to debate the exit sign.
The truth is that we've accelerated code generation without evolving how we verify it or who owns it.
Teams are producing more than they can reason about. Better abstractions, testing strategies, etc. aren't optional improvements anymore. Without them, none of this scales.
Finished doesn't mean closed.
One of my favorite things about the Multiplayer web dashboard is that you can reopen any session and keep going in the same thread the agent used the first time.
A bloated review queue full of AI-generated PRs isn't a sign that your team is moving fast.
Those who are addressing this problem are taking a step back and thinking about what data agents work from, how duplicate issues get caught, what makes a PR worth merging.
If you’re using coding agents, your output numbers (epics completed, task throughput, PR merge rate, …) may look good right now.
But those are not the metrics your manager is watching. Every single one of those numbers comes with an asterisk: production quality is declining.
Developers don't want more dashboards to stare at or more alerts to triage.
They want the thing that broke, to fix itself.
So we don't have to go through this flow of 👇 five tools, eleven steps and four hours just to fix a bug.
The “10x engineer” was always a simplification and I was very pleased when this term’s usage started to fade.
Then AI coding tools arrived, and it came roaring back, with even bigger numbers attached.
These look like 3 different problems but they share a common root: AI tools working with incorrect, sampled, or missing runtime data.
The teams making progress are asking harder questions about how they collect, correlate, and deduplicate data before it ever reaches an agent.
