Root and package operations
Privilege changes, host scripts, service replacement, package installs and persistent system configuration.
A confirmation message inside a chat is not the same as an enforced approval. A meaningful approval workflow captures the exact proposed action, sends it to a separate trusted review path, applies a scoped decision and records what happened.
An AI agent approval workflow is an authorization protocol: a bounded agent proposes an operation; a trusted policy/manager checks the request; a human approves or rejects sensitive changes; only the permitted action executes, and the outcome is logged. It should not be possible to turn the approval off by changing a prompt.
Privilege changes, host scripts, service replacement, package installs and persistent system configuration.
Removing messages, modifying protected records, mass edits, database mutations and broad filesystem writes.
Posting to production websites, sending external emails, disclosing confidential information and spending funds.
New credentials, scopes, network access, skill installations, shared workspace mounts and privileged devices.
The agent names the intended operation and target; arguments are validated and associated with its identity. Arbitrary prose alone does not authorize execution.
The manager checks whether the operation is permitted under the saved policy, allocated resources and allowed targets. Deny unsupported or unclear requests rather than interpreting them optimistically.
The operator sees purpose, exact proposed commands, likely side effects, requested privilege, and alternative responses. Absence of a response is not consent.
Bind approval to the proposed operation and verify results. A changed script, target or authority level should need new review.
Store decisions and execution logs durably, reconcile interrupted work, and report partial failure instead of claiming success prematurely.
XMAN 0.4.76a1 source describes special requests in a durable operator inbox and a daemon-mediated manager path, along with resource-policy choices. This is a development alpha; each deployment must verify actual socket permissions, process ownership, sandbox boundaries, host-script authorization and service-restart behavior.
No. They can prevent untrusted instructions from directly causing certain high-risk actions if the boundary is correctly enforced. Prompt injection can still influence an agent's analysis or proposed request, and an operator can still approve a harmful change. This is why skill review, least privilege, authentication, network policy, context separation and human judgment also matter.
Four concise paths through the architecture—written for operators, developers, security teams and evaluators. Each distinguishes configured controls from goals that still need testing.
Firsthand issue reports, exact failure symptoms, the proposed engineering response and explicit verification limits.