Stacked diffs were a solved problem in the early 2010s but were only relevant for monorepos with hundreds or thousands of engineers (Facebook, Uber, Twitter, Google).
With agent-driven coding, small teams now face the same constraints. We should expect the same across the stack
- Production engineering agents are bottlenecked by context. A single trace in a large distributed system can exhaust the entire context window. At Antimetal, we built a set of compression techniques that fit 10–100× more telemetry into our agents’ context windows, allowing them
- We're renting out a theater to prescreen The Odyssey two days before it's officially out. An epic about endurance, judgment under pressure, fighting monsters, and getting the team home in one piece...AKA a day in the life of a production engineer. July 15th in NYC
- The rarest skill right now is mono-tasking. Everyone can do 10 things at once. Very few can do just one.
- We ran an experiment. We pointed Antimetal at our own production environment, emitting 3,000 error logs a day, with no prior guidance. It cut our error volume by 76% in a single day. It found six unique problems, including one we'd misdiagnosed for weeks. It diagnosed and


