An agent that can work for 20 hours does not merely need a larger context window. It needs a plan for which model thinks, which model executes and how the system notices when the work begins to drift.
Runpod and Moonshot AI's Kimi in Practice webinar brings that operating problem into a 30-minute session on November 12. The program runs twice, at 9:00 a.m. and 2:00 p.m. Eastern, for developers and ML engineers building agents or coding workflows on open models.
Tang Feihu, Senior R&D Engineer and Developer Relations Lead at Moonshot AI, joins Runpod Forward Deployed Engineer Chris Vela. They will show how teams can use Kimi K3.1 as the lead agent while K2-series models handle subagent work, then connect the Kimi CLI to models running through Runpod Public Endpoints.
Long-running agent systems turn model choice into an operating decision. The most capable model does not have to perform every step. A system can reserve deeper reasoning for the moments that need it, move supporting work to other models and reduce the token cost of keeping the task alive.
That division creates another responsibility. A workflow that runs for hours needs more than a prompt. It needs delegation, checkpoints and enough control to keep one wrong turn from carrying through the rest of the job.
The webinar will cover how to divide work between K3.1 and K2 subagents, reduce token use without giving up output quality and set up long-horizon tasks. The speakers will also demonstrate a game built with Kimi, Blender and Unreal Engine and preview what comes next for Kimi.
Registration and session access are controlled by Runpod. Use the official page for the current registration details.
Thursday, November 12, 2026, with sessions at 9:00 a.m. and 2:00 p.m. Eastern.
Tang Feihu of Moonshot AI and Chris Vela of Runpod.
Developers and ML engineers building agents or coding workflows on open models.