20 hour agent execution benchmark

A 20 hour task budget is a signal that Proximal is aiming at the part of coding work where current agents usually fail, not on writing a small patch, but on staying coherent through a full day of research, debugging, testing, and redesign. FrontierSWE was built around problems like compiler optimization, ML optimizer design, and database reimplementation, where progress depends on many linked decisions and partial credit matters more than a one shot pass or fail score.1 2 3

• This is much longer than older public coding benchmarks. FrontierSWE describes SWE-Bench Pro solutions as averaging 107 lines of code, and says many Terminal-Bench runs last 1 to 20 minutes, while its own tasks give agents 20 hours because the work is much larger in scope.1
• The benchmark design is meant to measure endurance, not just correctness. FrontierSWE grades partial progress on a zero to one scale, and v2 says the harness keeps agents aware of remaining time so they keep iterating instead of stopping early. That makes the benchmark closer to real engineering loops of test, inspect, revise, and rerun.1 2
• The task mix maps closely to Proximal's product thesis. The company says it targets data for ultra long horizon software work, including legacy system modernization, and FrontierSWE includes outside collaborators like Modular, Prime Intellect, and Thoughtful Lab to source hard real world tasks that resemble those environments.1 4 5

This points toward a market where model quality is judged less by benchmark saturation and more by how long an agent can keep making useful decisions inside messy technical systems. As that becomes the standard, the valuable layer shifts toward the data, harnesses, and evaluation setups that teach and measure long horizon execution.2 3 4