Notion Lets Agents Run for 24 Hours and Meters Them, Copilot Studio Agents Get Their Own Mailboxes, and Google Previews Gemini for Government

The chatbot is quietly turning into a coworker, and this week's launches show what that costs: runs that last a day instead of minutes, agents with an identity of their own, and governance sold as part of the product. Details below come from vendor announcements and roadmap entries, so treat timing and performance claims as the vendors' own.

1. Notion Rebuilt Its Agents for 24-Hour Runs, and Started Metering Them

On October 19, Notion said it rebuilt its agent harness so Notion Agents can work on a task for up to 24 hours. Co-founder Akshay Kothari said the limit replaces a five-minute cap. Agents also get their own computer and browser, better context management, and a "Review for me" mode with safety checks.

The less glamorous half is the economics. Notion introduced a credit system, with admins able to buy extra credits now. A usage dashboard and spending limits (an "AI Command Center") are still listed as coming soon. It also added an auto model router that picks a model per task. Notion says that across 20-plus common tasks the router delivered results similar to Opus 5.5 and Astra at roughly 90% lower average cost. That is a vendor benchmark with no published methodology, so verify it on your own workload.

The lesson for builders: once a run can last a day, a bug is a billing event. Per-run budgets, step limits and a kill switch stop being nice-to-haves.

2. Copilot Studio Agents Are Getting Identities of Their Own

A Microsoft 365 roadmap entry (RM570430) describes persistent identity for Copilot Studio agents: each agent gets its own mailbox and Teams presence plus Office access, and admins govern it as a first-class entity. The entry lists a September preview and a November general-availability target. It is a roadmap item, so dates can move, and I could not confirm the preview is live.

This follows a pattern from recent weeks: agents that act under a borrowed human login are hard to audit, so vendors are moving to separate, revocable identities. If your agent sends email or joins meetings, a service identity with scoped permissions and its own audit trail beats a shared token.

3. Google Previews Gemini for Government

On October 20, Google Cloud opened a preview of Gemini for government services for U.S. federal, state and local agencies. It bundles reusable workflow skills, secure connectors to agency systems, partner agents and centralized governance. Google named the State of Missouri and the U.S. Department of Health and Human Services as early users. It follows the October 8 launch of Gemini's general work agent, which put finance and legal variants in preview. Coverage so far is press summaries, and how it performs across messy legacy systems is an open question.

What Connects the Three

Long runs, distinct identities and governance layers are the same shift seen from three angles: agents are being treated as operated infrastructure, not features. A rough checklist for your own deployment:

const agentRun = {
  identity: 'svc-support-agent',   // its own principal, not a user's token
  maxDurationMs: 24 * 60 * 60_000, // hard ceiling
  maxCredits: 500,                 // stop and ask a human past this
  requireReviewFor: ['send_email', 'delete', 'payment'],
};

A cost-per-run chart across a router's model choices would make a strong visual for a follow-up post.

Conclusion

The question for chatbot teams is no longer whether an agent can do the task. It is who it acts as, how long it may run, what it may spend, and who reviews it before it acts.

— Maya