Agents that provision accounts need identities of their own, Parallels' product chief argues

Task-scoped permissions move the constraint from model choice to the authorization queue. In the example Koryakina uses, one agent onboarding a new employee retrieves HR information, creates accounts, provisions applications and notifies managers ^[8]. The workflow touches four systems ^[18]. Traditional identity systems were designed around human users, who authenticate and receive permissions based on their role ^[13], so the cheap way to make that agent work is to run it as the employee who sponsored it.

Koryakina wrote that "Allowing an agent to simply inherit a user's credentials creates unnecessary risk" ^[6]. Her reason is proportionality: an agent may need only a small subset of the resources available to the individual it represents ^[17]. The alternative she sets out is delegated identity, where an agent receives only the permissions a specific task requires, and the organization can then determine who authorized it and whether its actions remained within policy ^[7].

She frames the whole deployment question as four items: who authorized the agent, what systems it can access, what actions it is permitted to perform, and how those actions are monitored and audited ^[9]. Enterprise identity, she wrote, will increasingly need to encompass both people and the AI agents acting on their behalf ^[14].

Some of that is available this quarter and some is not. Assigning agents their own identities with task-specific permissions that can be monitored and audited is a configuration decision now ^[15], as is a central approval process for AI tools and integrations of the kind many companies already run for SaaS ^[11]. The registry layer is later. Model Context Protocol is emerging as a standard for connecting AI systems to enterprise applications and data, and the column puts approved MCP registries no higher than "may become an important governance layer" ^[10]. Orchestration platforms that coordinate identities, policies, applications and governance are for the company running hundreds of specialized agents ^[12].

Koryakina is Chief Product and Technology Officer at Parallels, leading product strategy, engineering, DevOps, AI and infrastructure ^[1], and the argument she makes is that the binding problem sits in infrastructure. Two of the claims come from elsewhere. Microsoft's purpose-built identities for AI agents are another company's product decision ^[5], and Deloitte's 2026 State of AI in the Enterprise research is a survey finding ^[3]. The column recommends this work without attaching a cost or a timeline to any of it.

The 21% is self-reported maturity of a governance model, not evidence that identity is what stalls an agent programme ^[3]. By subtraction, 79% of the organizations Deloitte asked did not have a mature model ^[4]. Against that, the sequencing claim is the one I would act on: Koryakina wrote that organizations should establish governance, security and architecture standards before individual teams begin deploying agents at scale ^[16]. Do it in the other order and the correction is re-permissioning every agent already in production, along with every tool and data connection each of them reached ^[11].