AI Agents Run for Days: Orchestration Breaks

The landscape of AI agents is shifting, with new models demonstrating the ability to run continuously for hours and even days. Moonshot AI's Kimi K2.6 is engineered for this long-horizon execution, supporting autonomous tasks such as incident response and system monitoring. One internal use case saw an agent operate non-stop for five days.

This evolution highlights a significant challenge in AI orchestration. Traditional frameworks, built for shorter agent lifecycles, are struggling to cope with the stateful demands of continuous, long-duration tasks. Competitors like Anthropic and OpenAI have introduced multi-session and background execution, but Kimi K2.6's approach relies on the model itself to dictate orchestration rather than pre-defined roles.

Concerns are rising about the governance and management of these advanced agents. As they can generate code and system changes rapidly, organizations face challenges in review, remediation, and oversight. The architectural shift from scripts to agents as persistent infrastructure necessitates new management paradigms, including agent runtimes and mesh systems.