Long-horizon agents at half the cost (unreallabs.ai)

Unreal Labs announced significant improvements to their Unreal Agent harness with the introduction of a context compaction mechanism, which allows the agent to perform long-running tasks at nearly half the cost of its competitor, Codex, while achieving identical performance metrics. Utilizing GPT-6.1 Sol with xhigh reasoning, this new mechanism smartly summarizes the session context without disrupting ongoing tool calls, thereby maintaining its seamless asynchronous operation. This innovation not only enhances efficiency by reducing the total inference cost to $445.75 compared to Codex's $850.52, but it also ensures that the agent processes tasks more effectively. The significance of this update lies in its potential to lower operational costs for AI applications that require long-horizon decision-making while preserving task efficiency and execution speed. Additionally, Unreal Agent has launched a terminal user interface (TUI) to simplify user interaction and facilitate testing without needing extensive SDK knowledge. The integration of Claude models via Anthropic’s Messages API further highlights the versatility and adaptability of Unreal Agent, showcasing its capability to handle stricter messaging and content delivery requirements, which is crucial for maintaining asynchronous call integrity in complex tasks.