A 232× kernel experiment makes verification the agent skill
A Codex-driven kernel experiment reported a 232× speedup. Its useful lesson is not unsupervised generation: benchmarks, profilers, and a verifier could tell the agent when it was wrong.
The experiment paired an agent with benchmarking, profiling, research, and correctness verification. That made optimization a constrained search problem.
HN readers reported that agents work best where a system supplies a hard answer. They fail more readily when success depends on credentials or a human-operated dashboard.
A loop with an answer key
The auto-research write-up describes a benchmark, profile, verify, research, and improve cycle for a kernel. The reported result matters less than the loop: every candidate change is judged by a measurable outcome.
HN supplied the boundary
In the thread, a reader described an agent improving a video codec with compiler profiling, VTune, and a bitstream verifier. Another audited ten agent-completed tasks and found that the failures were tasks whose result was not machine-checkable.
Give automation an oracle
Start agents where tests, benchmarks, parsers, or simulators can catch regressions. When success is social, proprietary, or ambiguous, the correct design is a human handoff.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
Working with agents feels like leadership, until it doesn't
A developer proposed that working with AI resembles leadership. HN agreed that clear delegation matters, but argued that agents lack the agency that makes people management difficult.