Orchestration
A Regulated Asset Manager Is Holding Back Agent-to-Agent Communication on Purpose
Okta's Oktane 2026 event surfaced a detail that cuts against the usual agentic AI pitch: one of its customers, a large regulated asset manager, is deliberately throttling communication between its own agents while it works out how to govern it.
The same company's discovery scan found roughly 13,000 agents running in its environment, of which it considered only about 1,000 valid. It also found more than 500 MCP servers against a handful that had been formally approved. Those numbers describe a sprawl problem: agents and tool-serving infrastructure accumulating faster than any registry or approval process can track.
What's more notable is what the company did next. Rather than opening up agent-to-agent handoffs across that sprawl, it applied differentiated, narrow permissions at the individual agent level — a processing agent gets read-only access, another agent is permitted to update a record, even when the human behind both agents has broader standing rights. Updates can be time-bound rather than persistent. And the firm is extending that same restraint to agent chains: it is holding back broader agent-to-agent communication specifically until it finishes a reference architecture for governing those interactions.
That's a deliberate sequencing choice — discovery and narrow per-agent authorization first, chained autonomy later, gated on having an actual governance model rather than shipping orchestration first and patching accountability in after the fact. It runs counter to the more common pattern this year, where orchestration and handoff capability ships ahead of any accountability layer and the audit trail gets bolted on afterward.
It also implies something about where the real risk sits. The firm isn't worried about any single agent doing something wrong in isolation — fine-grained, time-bound, least-privilege access already addresses that. The open problem is what happens when authority passes from one agent to another, each potentially acting on a different slice of permissions than the human nominally responsible for the chain. Without a reference architecture for that handoff — who's accountable when agent A's limited authority combines with agent B's different limited authority to produce an outcome neither was individually permitted to cause — the company has concluded it's safer not to allow the chain at all yet.
This is a rare example of a production environment treating agent orchestration as the harder, later problem rather than the headline feature. Most vendor messaging around agent-to-agent protocols assumes chaining is the goal and governance catches up. Here, a regulated enterprise with actual exposure inverted that order, and did so only after a scan revealed that over 90% of its agent population was already operating outside its own inventory.