Agents & InferenceHacker News

Unreal Agent

Which summary reads better? Pick one — models revealed after.Both summaries are AI-generated.

Match the models (Optional)

Which model wrote which summary? Select a matchup mapping below before voting.

Summary A

An asynchronous tool-call harness cuts agent operating costs up to 40% versus Codex (20% vs Pi) with no measured performance loss, by logging tool calls as "in-progress" immediately and firing the LLM only when results return—so the model never burns tokens polling, waiting, or managing heartbeats. This means agents can run long tasks (multi-minute env setup) in parallel with exploration, accept user steering mid-flight without blocking, and shift security/approvals out of fragile harness hooks into deterministic sandbox constraints. The catch: it's a proprietary harness you'd adopt rather than an open SDK, so evaluate the cache-preservation and lifecycle claims against your existing provider integrations before committing.

LinkedIn

Two AI summaries of each story, blind-voted — see today's agents & inference digest →