There are 9 control objectives here. Each one states what it asks for, the risk it addresses and how it is measured. It then says what each risk tier requires, what evidence proves it, and who may check that evidence. The Operating Model page explains the tiers.
DSN-01Architecture standards
Design agents to the organization's architecture standards and patterns, so that they can be maintained and can work with each other.
- Risk addressed
- Inconsistent designs that are hard to maintain
- Key metric
- Share of agents that conform to the architecture standards
What each tier requires
- Tier 1 Exploratory
Experiments may use any pattern inside the sandbox.
- Tier 2 Operational
The design uses approved patterns, or records why it does not.
- Tier 3 Critical
An architect reviews conformance before build starts.
- Tier 4 Strategic
The architecture board approves the design and any exception.
- Evidence
- The design record, the patterns it uses, and the review decision.
- Who may check it
- Internal audit.
Scope each agent to one job it can do well, so that designs do not grow complicated and drift beyond their purpose.
- Risk addressed
- Agents that grow beyond their purpose
- Key metric
- Average number of declared offerings per agent
What each tier requires
- Tier 1 Exploratory
No limit applies in a sandbox.
- Tier 2 Operational
The design states the one job the agent does.
- Tier 3 Critical
An architect challenges any agent that declares more than one job.
- Tier 4 Strategic
Any agent that combines jobs needs a written reason approved by the architecture board.
- Evidence
- The agent's declared offerings and the design review that accepted them.
- Who may check it
- Internal audit.
DSN-03Declared scope and refusals
Write down what each agent does and what it refuses to do, and sign that declaration, so that its scope is clear to anyone who relies on it.
- Risk addressed
- Agents that drift into other agents' work
- Key metric
- Share of agents with a signed declaration of what they do and refuse
What each tier requires
- Tier 1 Exploratory
The owner describes the experiment's scope in a sentence.
- Tier 2 Operational
The agent has a written declaration of its offerings and refusals.
- Tier 3 Critical
The declaration is signed and published with every version.
- Tier 4 Strategic
The declaration is checked by someone other than the builder, and conformance to it is tested at release.
- Evidence
- The signed declaration for the version that is running.
- Who may check it
- Internal audit and a partner.
Design agents so that each decision and act can be explained afterwards, from records of what the agent saw and did.
- Risk addressed
- Decisions nobody can explain
- Key metric
- Share of agents whose decisions can be traced step by step
What each tier requires
- Tier 1 Exploratory
The experiment keeps a log of its runs.
- Tier 2 Operational
Each act is recorded with its inputs and result.
- Tier 3 Critical
Every step, model call and tool call is traced, so a single run can be followed from start to finish.
- Tier 4 Strategic
Explanations can be produced for an affected person or a regulator on request, within a set time.
- Evidence
- Trace records for sampled runs, and the time taken to answer explanation requests.
- Who may check it
- Internal audit and an outside auditor.
Design agents to standard interfaces and protocols, so that they can work with existing systems and with other agents.
- Risk addressed
- Agents that cannot work with other systems
- Key metric
- Integration test pass rate
What each tier requires
- Tier 1 Exploratory
No requirement applies in a sandbox.
- Tier 2 Operational
The agent uses the organization's approved protocols for messages and tools.
- Tier 3 Critical
Integration with each declared dependency is tested before release.
- Tier 4 Strategic
Integration is tested with the partners and systems outside the organization that the agent will deal with.
- Evidence
- The declared interfaces and the integration test results for the current version.
- Who may check it
- Internal audit and a partner.
Review security at design time, so that the harm an agent could do is limited before it is given any credentials.
- Risk addressed
- Weaknesses built into the design
- Key metric
- Share of agents with a design-time security review
What each tier requires
- Tier 1 Exploratory
The experiment has no production credentials.
- Tier 2 Operational
A standard threat checklist is completed for the design.
- Tier 3 Critical
A security reviewer sets the limits on what the agent can reach and do before build starts.
- Tier 4 Strategic
The security review includes a threat model reviewed by a second security reviewer.
- Evidence
- The completed checklist or threat model, and the limits it set.
- Who may check it
- Internal audit and an outside auditor.
Review and approve the design of every agent at Tier 2 or above before development begins.
- Risk addressed
- Flawed designs reaching production
- Key metric
- Share of Tier 2 and higher agents with an approved design
What each tier requires
- Tier 1 Exploratory
No design review is required.
- Tier 2 Operational
A peer reviews the design.
- Tier 3 Critical
The architecture board reviews the design in full.
- Tier 4 Strategic
The architecture board reviews the design, and the executive sponsor approves it.
- Evidence
- The design review record with the reviewers and their decision.
- Who may check it
- Internal audit.
DSN-08Finding ungoverned agents
Look for agents running without governance, and bring them under it through a simple path that does not punish the people who built them.
- Risk addressed
- Ungoverned agents on production systems
- Key metric
- Ungoverned agents found each quarter
What each tier requires
- Tier 1 Exploratory
Builders are told how to register an experiment.
- Tier 2 Operational
Network and spending data are scanned each quarter for agent activity that is not in the catalog.
- Tier 3 Critical
Scans run each month, and an agent that is found is registered or stopped within a set time.
- Tier 4 Strategic
Scans run continuously, and unregistered agent traffic is blocked.
- Evidence
- The scan results, the agents found and what happened to each.
- Who may check it
- Internal audit.
DSN-09Sandboxes for experiments
Run experimental agents in isolated sandboxes, so that risky trials never touch production systems or data.
- Risk addressed
- Risky experiments on production systems
- Key metric
- Share of experimental agents in a sandbox
What each tier requires
- Tier 1 Exploratory
The experiment runs only in a sandbox with synthetic or public data.
- Tier 2 Operational
A production agent's new version is tried in a sandbox before release.
- Tier 3 Critical
Sandboxes for Tier 3 agents copy the production limits, so that a trial behaves as production would.
- Tier 4 Strategic
Sandbox results are kept as part of the release evidence.
- Evidence
- The sandbox's network and data settings, and the list of agents running in it.
- Who may check it
- Internal audit.