Organisational Roles as Contracts for AI Agents

[AlexMikhalev]

New position paper on reference-architecture.ai, continuing the argument David Hunt and I made in "AI Risk is evolving, every day."

The question: how do you supervise an agent that can pursue a legitimate objective through methods outside its mandate? The paper proposes organisational roles as contracts. A role defines the governed knowledge graph an agent uses, the capabilities it supplies through services, and the tool actions it may take under contextual permission.

What it covers:

• The chain from organisation to role, capability, service, tool, action and contextual permission
• Why concept coverage is not verification, and what a semantic compatibility contract between requirement, output and criteria needs
• How Terraphim role graphs, checklists and Terraph proof-gated verbs fit the contract
• A standards cross-check against ISO/IEC/IEEE 42010:2022 and 29148:2018
• An evaluation plan with expected observable results for each boundary case

Full paper: https://reference-architecture.ai/posts/organisational-roles-as-agent-contracts/
Short version: https://reference-architecture.ai/posts/from-ai-risk-to-accountable-agents/

Feedback welcome, especially on typed relationship profiles and how you bind role assignments to execution.