Blog ·
Management: AI trends to focus on - governing persistent agents across teams
Management focus is shifting from picking AI tools to governing persistent agents that work across teams. You now need to define who owns each agent, what it can decide, and how to measure failures.
Jeroen Erne
This week the management conversation shifted from picking AI tools to governing persistent agents that operate across teams. The practical question is no longer “which model” but “who owns the agent, what can it decide, and how do we measure when it fails.”
What changed this week
Agentic AI moved decisively into shared team workflows. xAI launched Team Bots designed for group workspaces, Meta rolled out an enterprise AI platform with a new initiative leader, and OpenAI released Dots, always-on agentic avatars that persist across sessions. These are not single-user assistants. They sit inside channels, documents, and recurring processes where multiple people interact with them.
Infrastructure for controlling these agents arrived in parallel. Nvidia announced a full-stack platform for reining in rogue agents, OpenClaw shipped an enterprise control plane for persistent agents, and Restate raised $20 million for durable agent infrastructure that survives failures. Cloudflare opened a waitlist for managed enterprise agent workspaces. The message is clear: agents are becoming operational, and the control layer is being built underneath them.
On the model side, Anthropic released Sonnet 5.5 as a cheaper, faster work model, while DeepSeek Harness added desktop apps, plugins, and scheduled automation. Manus 2.0 introduced persistent computers and event-triggered agents. The cost and capability curves are both bending toward always-on deployment.
Management attention is also shifting toward metrics. Boards are being urged to track stakeholder momentum around AI transformation. The consumer AI market showed the strain of costly inference and uncertain retention, a warning for enterprise teams scaling agent usage without usage governance.
What it means for you
You are no longer just selecting AI tools for individuals. You are designing how shared agents enter team workflows, what they can access, and who answers when something goes wrong. The products launched this week assume agents will run continuously across documents, apps, and operational processes. Without explicit decision rights, escalation paths, and failure metrics, you are deploying risk, not productivity.
Your role is becoming queue supervision, cost enforcement, and permission design. The new control planes from Nvidia, OpenClaw, and Cloudflare exist because agents drift, exceed bounds, and generate work that looks complete but is wrong. You need to make agent status visible to your team, enforce budget and approval boundaries, and protect human review capacity. This is not about blocking automation. It is about making it accountable.
Metrics matter differently now. The week’s stories point toward measuring business outcomes rather than activity volume. If you reward agent usage counts, you will get usage. If you measure resolution quality, review workload, and exception rates, you will get governed adoption. Cohere’s new retrieval metric, focused on rank consistency, is a reminder that even technical measures are shifting toward reliability over speed.
Your team structure will change. When agents become persistent participants, roles shift toward collaboration between people and automated workers. The youth job program adding fact-checking and prompting to workplace readiness is an early signal: the skills you hire for and develop are changing. Explicit owners for each agent, visible work histories, and exception queues are not nice-to-have. They are the scaffolding that lets you scale without losing control.
What to focus on next week
- Pick one recurring team workflow and define the decision rights an agent would need before you turn anything on. Write down what it can decide, what it must escalate, and who owns the outcome.
- Audit your current AI tool spend and usage. If you cannot see cost per team, failure rates, or review workload, flag that gap before expanding access to persistent agents.
- Identify one metric tied to a business outcome—resolution time, rework rate, customer effort—that you will use to judge agent performance instead of counting prompts or sessions.
- Schedule a 30-minute discussion with your team about what work they would stop doing if an agent handled a defined portion of it. Use that to set boundaries, not to promise headcount reduction.
- Review the control and visibility features in any agent product you are piloting. If you cannot see work history, set spend limits, or force human approval at defined points, treat it as experimental only.
These recommendations draw on a full week of product launches, funding announcements, and market signals. For the complete set of stories, see all Management AI news.