Agents that run for days are cracking the orchestration stack
Moonshot's Kimi K2.6 coordinates 300 subagents across 4,000 steps and has run incident response for five straight days. The frameworks underneath were built for tasks that finish in seconds.
- Most orchestration frameworks were built for agents that run for seconds or minutes, and they are starting to crack as production agents run for hours or days, a VentureBeat analysis argues.
- Moonshot AI's Kimi K2.6 manages up to 300 subagents across 4,000 coordinated steps, and the company has run agents for five straight days on monitoring and incident response.
- The unsolved problem is state: keeping a long-horizon agent coherent while the environment changes mid-task, and keeping observability stacks calibrated for human-paced traffic from going blind.
The agent industry spent two years optimizing for tasks that finish before the demo ends. The next fight is duration. A VentureBeat analysis circulating this weekend argues that Moonshot AI's Kimi K2.6, which can run agents for days without human intervention, exposes a structural gap: most enterprise orchestration frameworks assume agents live for seconds or minutes, and they are starting to crack as runtimes stretch to hours and days.
Three hundred subagents, four thousand steps, five days
The K2.6 numbers set the scale of the problem. Its Agent Swarm capability manages up to 300 subagents executing across 4,000 coordinated steps, up from 100 subagents and 1,500 steps in the prior release, per MarkTechPost's launch coverage. Moonshot says one internal agent ran five straight days handling monitoring and incident response. Notably, K2.6 lets the model itself determine orchestration rather than assigning predefined roles, and its Claw Groups preview opens the swarm to external agents. The hard part is not spawning the swarm, it is state: a five-day agent must stay coherent while deploys, data, and permissions change underneath it mid-task.
Observability built for humans goes blind
The same weekend's practitioner chatter added the operational half: always-on agents are overwhelming enterprise observability stacks calibrated for human-paced traffic, creating production blind spots. That echoes what InformationWeek found surveying the tooling: no vendor currently watches every layer, and multi-agent systems fail in cascades, not single errors. The platform vendors are moving the orchestration primitives down-stack, as OpenAI did by building multi-agent into its API, but duration is a different axis than fan-out. An agent that works for five days is an employee-shaped workload running on infrastructure that still thinks it is serving a chat message.