AI coding agents should use their own narrowly scoped identities rather than borrow a human developer’s login, according to KloudStax CTO Vinay Thakker in a Help Net Security interview published November 2. For business tasks, the agent acts under its own identity; for work requested by an individual, the request also carries that person’s identity. The resulting action is limited by what both identities permit, while logs retain both identities.
Thakker’s example separates proposing an infrastructure change from applying it: an agent may open a change and run a plan, but application remains subject to the existing human review and deployment pipeline. He also recommends short-lived credentials obtained through a broker that the agent cannot read. The broker must still verify the agent, restrict destinations and avoid shared credentials, because credential brokering does not remove the agent’s authority.
The proposed coding-agent baseline extends those boundaries beyond the model. It includes a disposable environment restricted to one repository, approved build and test commands, blocked access to files such as .env and SSH keys, and network access limited to exact hosts. Package installs should use an internal artifact repository, while dependency additions, deployments, migrations and deletes require human approval. Thakker says branch protection, network restrictions and credential access should be enforced in version control, firewalls and identity systems respectively, so the controls remain in place even if an agent bypasses its own policy.
That enforcement boundary provides a concrete link to a separate April interview with AppViewX CEO Archit Lohokare. Lohokare described a unified control plane that would discover AI agents and machine identities, enforce governance, and monitor, audit and control their access and behavior. Read together, the interviews point to a deployment condition: an agent needs a distinct identity and task-level authorization, but those permissions must be connected to external systems that enforce network, credential, change-management and audit rules.
The evidence remains limited. Both documents are interviews published by Help Net Security with security-company executives; neither supplies measured results, an independently evaluated deployment or a production case study. Thakker’s recommendations therefore describe an adoption model, while Lohokare’s account includes claims about AppViewX’s platform rather than independent validation.