Supervision, by design.

RDES, being developed through the MCP server project, is intended to check that RDEP’s rules are being followed as a project progresses and as the accountable human works with agents. Its role includes matching the work to the relevant Gate, helping prepare the required project records and supporting the human’s review and authorization of the next step.

Identify the applicable Gate

The intended workflow relates questions, proposed work, findings and decisions to the project’s current stage and the relevant RDEP Gate. The applicable requirements and any missing evidence are made explicit, so progress is judged against the right criteria.

Prepare the work with the executing agent

Through dialogue with the human, the executing agent helps prepare and update the Charter and the other records required at that stage. RDES is intended to support this interaction, retain the sources and unresolved questions, and present proposed entries for human review and approval.

Prepare the inputs for Kernel evaluation

The Kernel is an optional evaluation tool. It reads project records and produces derived findings; it does not author the project’s authoritative records or grant human approval. In the intended RDES workflow, the executing agent and human prepare the Charter and the governing-configuration and decision records needed for an applicable, authorized evaluation. Proposed content is presented for human approval; the Kernel then evaluates the supplied material within its supported scope.

Review before passage

RDES is intended to check the applicable conditions, expose gaps and assemble the material for the corresponding Gate review. The accountable human examines that material and decides whether to approve passage, require further work or defer the decision. A completed form or a successful automated check is not a Gate decision.

Record approval and control the next step

The human’s decision must identify the material reviewed and the scope of the work authorized next. RDES is intended to record that decision and help keep subsequent agent activity within it. Human direction and the required approvals remain part of the interaction throughout the process, rather than a blanket permission inferred at the start.

The scope of supervision

RDES is intended to apply RDEP’s rules through connected human–agent workflows. For each supported integration, the scope must distinguish what information RDES receives, what it can assess, and which actions it can hold pending the required checks and, where applicable, human authorization. Activity outside that coverage must not be presented as checked or controlled.

RDES is being defined and designed as a supervisory product and interaction interface between humans and AI agents, with the primary purpose of putting RDEP’s principles and rules into practice throughout their work. RDEP provides the governing rules; RDES is being developed to apply them through a human–agent interaction interface. The interface is intended to make the work, its evidence and unknowns, and the decisions required from the human understandable and actionable. Records, source traceability and conversation transfer support that interaction.

The principal capabilities we are pursuing for RDES are:

• Applying RDEP during work: enforcing applicable rules and prerequisites within the connected workflow; checking roles and authority, exposing missing support and controlling actions within that scope.
• Human–agent interaction: presenting the work, evidence, unknowns and required decisions in a form the human can understand, with ways to question, direct and review the agents’ work.
• Gate-by-Gate control: relating the work and decisions arising in human–agent interaction to the applicable RDEP Gate and its requirements; identifying missing evidence and preparing the appropriate review. Gate passage requires the accountable human’s explicit decision on the reviewed material.
• Assisted project records: working with the executing agent and the human to prepare and update the Charter and other required project records. Proposed content is reviewed with the human, and approval is recorded for the identified version; filling a form does not pass a Gate. Where Kernel evaluation is applicable and authorized, its project inputs are prepared from those records; the Kernel evaluates rather than fills or approves them.
• Traceable knowledge and decisions: preserving connections among sources, findings, assumptions, requirements and decisions for examination and reconsideration.
• Preserving unresolved issues: keeping open questions, risks, pending decisions and unfinished work visible.
• Coordinating roles: supporting collaboration among executors, advisors and reviewers, with human direction and reserved human decisions.
• Project continuity: retaining authorized history, context, negative results and recoverable records, including appropriate local storage.
• Human-directed transfer: moving work between conversations and agents with the necessary context, and configuring roles, instructions, access, models and effort levels in environments that support them.
• Automating repetitive checks: reducing manual work in recording, reconciliation and preparing information for review.
• Supporting revisable decisions: keeping alternatives, selection rationale and information that may warrant reconsideration available.

Automation is an important part of the product's objective. Each capability's effectiveness and limits must be tested; mechanical checks, substantive evaluation and human decisions each have their own place.