KDCube
← AI Industry News

Hosted Agent Runtimes Get Price Tags: Kimi Ships, Microsoft Starts the Meter

Moonshot AI's Kimi Hosted Agents went live Nov 5 with a managed harness, per-hour sandbox billing, and 1–20 child-thread delegation, while Microsoft's Foundry Copilot harness starts charging Dec 1. Meanwhile Kong folded agent discovery, allowed-tool controls, and cost caps into Konnect — the agent control plane is movi...

AI Industry Dossier · Agent Runtimes

Highlights

  • Moonshot AI's Kimi Hosted Agents went live Nov 5 — a managed harness for long-running and async agent work, with sandbox runtime billed at ¥0.072/hour and multi-agent delegation of 1–20 child threads. (Unite.AI)
  • Microsoft's Foundry Copilot harness starts billing Dec 1: sandbox compute at $0.0994/vCPU-hour and $0.0118/GiB-hour, and during preview Microsoft — not you — owns the approval policy for its five built-in tools. (beri.net)
  • Kong expanded AI Management inside Konnect on Nov 5, pushing agent discovery, allowed-tool controls, observability, and cost caps onto the API gateway. (AI Agent Store)
  • Reminder for operators: GitHub's connect-once MCP connectors (26 services) are still in public preview across Copilot surfaces. (GitHub)

Key Signals

  1. Kimi turns its agent harness into a hosted product Nov 5, 2026

    Moonshot AI launched Kimi Hosted Agents with a web console, REST API, and kha CLI, storing full session history as events for automatic resumption after platform failures. Coordinator agents delegate to 1–20 child threads (no nested children), and an enterprise self-host mode moves tool execution onto customer infrastructure while Kimi keeps inference and orchestration — though it drops IPython, image search, and save_artifact. (Unite.AI)

  2. Microsoft's free preview has a meter now Dec 1, 2026 start

    The Foundry Copilot harness stops being free Dec 1, with compute priced separately from model and tool charges. The sharper operator detail: custom approval logic applies only to MCP tools you configure, not the five built-ins (file read/write, shell, web fetch, subagents), and harness selection is permanent per agent during preview. Microsoft's own guidance warns that disabling file ops while leaving shell on "isn't a strict security boundary." (beri.net)

  3. Gateway vendors claim the agent-governance layer Nov 5, 2026

    Kong's AI Management push in Konnect — discovery, security controls, observability, and cost management for agent workflows — lands the same week as CrowdStrike's agent insider-threat detection and Kong-style budgets/logging thresholds. The control plane is consolidating onto infrastructure teams already run. (AI Agent Store)

Why It Matters / What To Watch

  1. Hosted runtimes are now a priced, comparable market — read the fine print.
    • Compare Kimi's ¥0.072/hr sandbox against Microsoft's $0.0994/vCPU-hr before December locks in your Foundry bill. (Unite.AI)
    • Get written answers on approval policy, idle billing, and switching paths before Microsoft's Nov 30 preview cutoff. (beri.net)
  2. Who owns tool approval is becoming the real differentiator.
    • If the platform controls built-in-tool approval (Foundry) while you only govern MCP tools, map which actions you can actually gate. (beri.net)
    • Watch gateway-side governance (Kong, connect-once connectors) as the place to enforce allowed tools and budgets across agents. (AI Agent Store, GitHub)

Quick Links