How Cerebras Keeps Autoresearch Loops From Going Off Track: Lessons From 71 Experiments
Cerebras had an AI agent run 71 experiments across two task types—training optimization and model compression—and found that when the loop scope is tightened, it reliably produces real conclusions, while once the guardrails are loosened, the agent drifts from the goal within hours. In experiment one, codex-autoresearch-harness was used to wrap Codex in a bash loop; GPT-5.4 and Codex-Spark each independently converged on a learning-rate warmdown schedule within 6 hours, but GPT-5.4 had a proposal acceptance rate of 67% versus only 17% for Spark, showing that proposal quality determines total cost more than single-inference speed. In experiment two, REAP plus INT4 quantization compressed a 717 GB model to 92 GB, with dynamic expert swap-in/swap-out taking 0.151 seconds, but after 12 hours of unattended operation the agent changed to asking a different question on its own; the authors corrected this by clearing context, isolating directories, and adding more frequent validation checkpoints.
Cerebras had an AI agent run 71 experiments across two task types—training optimization and model compression—and found that when the loop scope is tightened, it reliably produces real conclusions, while once the guardrails are loosened, the agent drifts from the goal within hours. In experiment one, codex-autoresearch-harness was used to wrap Codex in a bash loop; GPT-5.4 and Codex-Spark each independently converged on a learning-rate warmdown schedule within 6 hours, but GPT-5.4 had a proposal
Matched fieldsSummary、Body
