An AI agent should be able to prepare work and take authorised steps. That requires more than a language model: a role, relevant information, approved tools and a responsible person.
A personal assistant supports its associated employee. It works within the same access permissions and can use approved documents, tasks and prior work context. Being an AI does not grant broader visibility.
Communication with colleagues remains between people. Another employee’s personal assistant is not a back door to their information or decision-making authority.
This diagram explains the relationship. It is not a live view or a commitment to future features.
An independent agent is not assigned as one individual’s personal counterpart. It can handle a bounded task on behalf of management or an authorised person, such as preparing information, structuring a case or supporting an approved process.
It still needs ownership, permissions and an actual assignment. “Independent” does not mean it can grant itself arbitrary access, tasks or approvals.
The execution environment, or agent harness, connects context, planning and tools. It can include configured interfaces, file work or code execution in isolated environments. Permitted tools are defined for the intended use.
An assignment can be broken into steps, producing drafts, documents or proposed changes for review. Delegating subtasks does not expand the original authority.
The language model is one part of the system. Brezel holds the operational context of roles, records, cases and approved knowledge. Each task receives relevant information within the appropriate access scope.
The architecture includes supplying context again for each task. Events and configured schedules can trigger work. This does not imply unlimited parallelism, guaranteed continuous availability or error-free execution.
People define objectives, review important results and provide required approvals. Critical steps such as external commitments or sensitive record changes need deliberate treatment. Professional responsibility is not handed over to the model.
Automatic job-role templates, a fully developed role-based learning hierarchy and agents making phone calls are separate development topics. Today’s service and future extensions remain distinct.
AI support for real tasks. Within your rules. With people remaining responsible.
brezel