It's not just "who" anymore — it's "which agent"
Access control design used to be a two-category problem: human users and service accounts. That assumption is breaking down in 2026.
Autonomous agents embedded in production systems now call APIs, touch databases, and delegate tasks to other agents without waiting for human approval at every step. As this pattern has become common, some estimates put non-human identities — service accounts, workloads, and now AI agents — at roughly 50 times the number of human accounts inside the average enterprise. This post looks at what webhani considers when agent identity comes up in a client system design.
Why the old IAM model doesn't hold
Traditional identity and access management rests on a few assumptions:
- Identities are stable over long periods (an employee account persists for years)
- Delegation is explicit and traceable
- "Who did what, and when" maps cleanly to a single identity
AI agents break every one of these. An agent might spin up and tear down in seconds or hours. One agent delegates a task to another, which calls a third tool — a chain of delegation that makes "under whose authority did this happen" genuinely hard to answer after the fact.
Four architecture patterns taking hold in 2026
Current agent security implementations mostly fall into four patterns.
1. User-delegated
The agent acts as the user, under explicit user consent.
2. Autonomous
The agent holds its own standing identity and is authorized directly.
3. Hybrid orchestrated
One agent delegates tasks to other agents.
4. Scoped impersonation
The agent acts on the user's behalf, but only within a defined scope.Which pattern fits depends on what the agent is acting on behalf of, and how much authority it needs. The real risk is mixing patterns without a clear boundary — at audit time, you can no longer tell which permission model actually governed a given action.
What to get right when implementing this
When webhani designs agent integrations for clients, these are the first questions we ask.
1. Treat the agent as a first-class identity
Don't model an agent's service account as an extension of a human account. Give it its own lifecycle, scope, and audit trail.
# Conceptual example: minimal agent identity shape
agent_identity:
id: "agent-invoice-processor-01"
type: "autonomous"
scope:
- "read:invoices"
- "write:payment-status"
lifespan: "ephemeral" # expires when the session ends
delegated_by: null # autonomous — no delegator
audit_trail: true2. Make delegation chains visible
When agent A hands a task to agent B, that relationship has to be logged, or you lose the ability to reconstruct an incident later. Every delegation should record who granted what scope, to whom, and until when.
3. Keep scopes minimal and time-bound
With human API keys, "grant broad access now, tighten it later" is sometimes tolerable. Agents act at a much higher frequency and speed than humans, so over-provisioned scope is a proportionally bigger risk. Short-lived tokens paired with task-scoped permissions should be the default, not the exception.
Where webhani lands on this
Agent identity management is still a fragmented space — implementations differ across vendors, from native platform features like Microsoft Entra Agent ID to specialized IAM brokers like Aembit. But the underlying principles hold regardless of vendor: don't treat agents as an extension of human identity, make delegation chains traceable, and keep permissions short-lived and minimal.
Before adopting a new vendor product, webhani recommends first mapping which architecture pattern your current agent usage actually falls into. Choosing a product before that mapping is clear tends to force a permission-model rebuild down the line.
Takeaway
As agent adoption grows, the ability to trace "who actually did this" becomes more important, not less. Agent identity design has a mix of decisions you can fix later and decisions that are expensive to retrofit. webhani recommends building identity and access design into the first phase of any agent integration project, alongside the feature work itself.
Sources: What's shaping the AI agent security market in 2026 (Palo Alto Networks), 2026 IAM Vendors: Comparing Top Agentic Security Firms (Aembit)