We’re sharing Unreal Agent — an agent harness that delivers up to 40% cost savings compared to Codex on production workloads and coding/science benchmarks, without any negative performance impact.
Read in full here:
Interesting cost angle. A lot of harness savings come from fewer / shorter turns, but if the agent still resends a fat diary of tool logs and overlapping context every decision, the input meter keeps eating the win.
I’ve been treating that as a separate Tier-1 step: deterministic compress on the message/tool list before serialize (no extra LLM judge). Contextpress does that locally with an agent profile that keeps tool call/result pairs intact — useful when you’re already optimizing the harness.
pip install contextpress