Agentic AI is scaling faster than the governance to run it

Agentic AI has crossed from novelty to roadmap item. Across Hong Kong enterprises, agents are starting to do real work — drafting, triaging, retrieving, and increasingly acting across systems. Deloitte's 2026 State of AI report expects agentic usage to rise sharply over the year.

The capability is arriving faster than the discipline to run it. In the same research, only about one in five companies report mature governance for agentic AI.

For a CIO or CTO, that gap is the whole story. Launching an agent is now easy. Running a fleet of them — safely, observably, accountably — is not.

Agents aren't applications

The instinct is to treat an agent like any other feature: build it, test it, ship it. But agents behave differently from the software your governance was designed for.

They act, and they chain. An agent doesn't just return an answer; it calls tools, triggers workflows, and makes decisions across systems. A small misjudgement can propagate downstream before anyone notices.

They drift. An agent that performed well in March can quietly degrade by June as upstream foundation models update, prompts lose specificity, or the underlying data shifts. This behavioural drift is invisible to traditional monitoring — application performance tools watch infrastructure, not judgment.

They're hard to see into. Without a control point in front of agent traffic, there's no consistent way to enforce policy, verify identity, audit actions, or cap token spend across a growing fleet.

Physical AI raises the stakes further. Deloitte reports that more than 58% of companies already use some form of physical AI — cobots, inspection drones, autonomous forklifts — with Asia-Pacific leading early adoption. In Hong Kong's logistics, manufacturing and public-service settings, an agent's decision can now move something in the real world.

The shift: operate agents, don't just deploy them

The organisations pulling ahead treat agents as systems they run, not features they ship. Three capabilities separate them.

A control plane in front of agent traffic. One place to enforce policy, verify identity, route to the right model, audit every action, and govern token spend — put in place before agents scale, not after an incident. This is the discipline ASTRA Gate was built around.

Continuous stewardship. Agents need monitoring for behavioural drift and a recalibration loop — observe, diagnose, calibrate, reinforce — across people, knowledge and agent health. We call this Agentic Stewardship, and it treats agent reliability as an uptime concern rather than an afterthought.

Human accountability. Every agent needs an owner — someone who reviews its decisions, defines escalation, and can retire it when it fails. Autonomy without accountability is just unmonitored risk.

What a technical leader can do this quarter

• Inventory where agents already act — including the shadow ones teams have quietly shipped.
• Put a control plane in front of agent traffic before you scale, not after an incident.
• Define drift signals and a recalibration cadence for every agent in production.
• Assign an owner and an escalation path to each agent; treat "no owner" as "not in production."
• Pilot physical AI in a bounded, observable environment before widening the blast radius.

None of this slows agentic AI down — it's what lets you speed it up safely. The teams that can run agents dependably will deploy more of them, sooner, and with less risk than the ones still treating each agent as a one-off.

That operating layer — the control plane, the stewardship, the accountability — is the work ASTRA does for Hong Kong enterprises putting agents into production.