Unreal Agent
6 hours ago
- Unreal Agent is a harness that reduces tool-management overhead, achieving up to 40% cost savings compared to Codex on benchmarks and production workloads without performance loss.
- It manages tool calls asynchronously, allowing users to steer the agent without waiting for tool completion and enabling more tool work per model turn.
- Motivation from production experience: existing SDKs (e.g., Claude's Agent SDK) have limitations like local session assumptions, heavy dependencies, and security issues; Unreal Agent addresses these.
- Architecture: immediately appends 'in-progress' event-log records for tool calls and final results when done, without breaking cache.
- Cost efficiency comes from minimal harness footprint, token-optimized tool results, and more tool work per model turn.
- Benchmarks (Terminal-Bench, SWE-Atlas, DeepSWE, ALE-CLI) show similar or better pass rates with significantly lower cost compared to Codex and Pi.
- The SDK includes a Go library, a runner executable, and a benchmark runner compatible with Harbor.