Pushing limits of harness engineering | Engineering Lead, Agents @ Hypha
New York City
Joined November 2025
- Tweets611
- Following289
- Followers122
- Likes972
Shoutout @useblacksmith for building CI infra. Their CPU performance on x86 runners and caching really helped to get CI running under 3 minutes.
OpenAI Decisions API has 30x the Jev's context. But you don't have to use it all. An 8k context budget was enough to classify financial documents at 98.25% F1—with 84% lower model cost than our 480k reference.
Multi-label document classification is a great use case for Decision models. OpenAI's Decisions API achieves 98.1% F1 versus the agent’s 93.5% at 1/7 of the cost and 28× lower p90 latency.
Decisions API as “auto mode” for a robot.
Image: glassware
Action: pick it up
Decision: denied, high chance of breaking it
This is the only way I can let a robot roam my apartment when I am at work
One good thing we got from vibecoding is optimized dev tooling. Now we deal with such a large amount of code that old tooling started to choke, so it got migrated to rust / go / etc. Look at TS 7, 10x faster typecheck with tsgo embedded
Robotics is currently at a state where Applied AI was 3 years ago.
- AI agents dont know how to build
- One guy online solved a similar problem
- Insights are in research papers
- Twitter is full of scripted demos
Best time for real engineers that dont just “/goal” their way
On screen awareness in the new Siri is pretty cool. I just asked Siri to create a grocery shopping list from imessage on my screen
Build plugins.
I said that publicly on June 4th at OpenAI’s event.
Latest YC demo day proved it.
Yes, those startups built plugins on top of existing harnesses, not the harnesses themselves.
Aged well, should have twitted it.
We use word “harness” too broadly.
After the YC demo day people keep saying “Building harnesses is the new moat”
But those startups didnt rebuild claude code - they extended it with plugins: skills & tools.
Plugin doesn’t sound as fancy, but its the right term here.
The craziest K8s dashboard I’ve ever seen @ExtendHQ
- Spaceships - nodes; the bigger the ship - the more pods it’s running
- Asteroids - jobs. When it descends - processing starts
- Black hole - queue. If job exceeds the number of retries - it goes into the void
Can I get a PR review to improve concurrency in Langfuse JS SDK? @hassiebpakzad @marliessophie @langfuse
github.com/langfuse/langfuse…
