MTS - Agent Harness
Aside, San Francisco, California
What you will do
- Polish our browser automation framework and memory system. You can make everything from the ground up, from the agent loop, compaction, memory to browser automation framework. Here’s an article how we did.
- Build an agent that uses your computer. Help Aside navigate browsers and desktop apps, understand what’s on screen, and complete tasks across them.
- Optimize our harness. Reduce token usage, work around bot detection, and experiment with heuristics that replace long prompt instructions with predictable behavior.
- Run real-world experiments. Build benchmarks and evaluation environments to test changes, catch regressions, and understand where the agent still fails on real tasks.
- Build with users. Talk to them, understand where the agent falls short in their work, and turn their feedback into shipped improvements.
What you need
- Use your brain. No benchmaxxing or /autoresearch. You read the code and prompts yourself, make changes line by line, and understand why each change helps.