KDCube
← AI Industry News

Agents Go Unattended and Multiplayer — MCP Drafts Their IDs

Snowflake's CoCo Automations now schedule unattended agent runs — but they inherit every role a user holds, and EXECUTE AGENT TASK is granted to PUBLIC by default. Meanwhile the new MCP roadmap puts agent identity front and center, and Salesforce's Slack Code makes multi-v...

Lead signals

Highlights

  • Snowflake moved CoCo (Cortex Code) Automations to preview on Aug 21, letting anyone schedule recurring, unattended agent runs in a managed sandbox — each run spawns an inspectable Cortex thread (Snowflake Docs).
  • The catch: those scheduled agents run under the user's default role plus all secondary roles, and EXECUTE AGENT TASK is granted to PUBLIC by default — so autonomy shipped ahead of least-privilege (THE D*AI*LY BRIEF).
  • The MCP roadmap (Aug 22) puts agent identity front and center: Workload Identity Federation and proof-of-possession instead of API keys, plus long-running "Tasks" and progressive discovery (MCP Blog).
  • Salesforce shipped Slack Code (Aug 20): tag Claude Code, Devin, Copilot, or Vercel's agent into a channel and the team reviews diffs and signs off before anything ships (Salesforce).

Key Signals

  1. Snowflake CoCo Automations reach preview — and expose a privilege gap Aug 21–22, 2026

    Automations turn a prompt into a recurring, unattended run that executes even with your terminal closed, available across AWS, Azure, and GCP commercial regions (Snowflake Docs). But each run's task session activates the user's default role and all secondary roles (DEFAULT_SECONDARY_ROLES = ('ALL')), and the ability to create them is granted to PUBLIC by default — meaning any user can spin up an agent that reaches everything they can, unattended (THE D*AI*LY BRIEF).

  2. MCP's new roadmap targets the exact wall builders keep hitting Aug 22, 2026

    The Core Maintainers' roadmap names five priorities: agentic messaging primitives (long-running Tasks, streamed results, mid-flight steering), HTTP-native transport unification, standardized agent identity via Workload Identity Federation and Demonstrating Proof of Possession, improved tool-result contracts with progressive discovery, and better SDK conformance (MCP Blog). It's the direct answer to the "MCP auth is the top unresolved gap" refrain — now with a concrete design direction rather than a warning.

  3. Slack Code makes agentic coding a reviewable team activity Aug 20, 2026

    Mention a coding agent — Claude Code, Devin, GitHub Copilot, or Vercel's — and it opens a project-specific channel where teammates watch the work, audit diffs as they're proposed, preview HTML output, leave feedback, and approve before merge (Salesforce). Multi-vendor support at launch means the human-approval checkpoint lives in the collaboration surface, not buried in a single vendor's IDE.

Why It Matters / What To Watch

  1. Unattended autonomy is arriving faster than the guardrails.
    • If you enable CoCo Automations, revoke the default EXECUTE AGENT TASK grant from PUBLIC, re-grant to named roles, and set DEFAULT_SECONDARY_ROLES = () so scheduled agents don't inherit every role a human holds (THE D*AI*LY BRIEF).
    • Treat "runs when your browser is closed" as a production system: each Cortex thread is your audit trail — plan who reviews them (Snowflake Docs).
  2. Identity, not model quality, is the next platform battleground.
    • Watch the MCP identity work (Workload Identity Federation, proof-of-possession): it's what will let you distinguish and authorize which agent acted, across vendors (MCP Blog).
    • As approval surfaces move into shared channels like Slack Code, standardize where sign-off happens so multi-vendor agents don't each carry their own review model (Salesforce).

Quick Links