identity agent
AI agents: provenance, tools, runs, budgets and approvals
Applies to: Identity and access. Who may do what: principals, sessions, roles, policies, relationships, keys and certificates, machines, devices, delegation and agents.
Written in commons/data/models/identity-agent/mod.ts:214. Requires identityDelegation, identityRebac. 8 concepts, 10 code lists.
Diagram source
flowchart LR
agent -->|principal| principal
agent -->|sponsor| principal
agent_budget -->|agent| agent
agent_budget -->|delegation| delegation
agent_budget -->|run| agentRun
agent_run -->|agent| agent
agent_run -->|delegation| delegation
agent_run -->|initiator| principal
agent_run -->|parentRun| agentRun
agent_run -->|tenant| tenant
agent_run -->|version| agentVersion
agent_tool -->|permission| permission
agent_tool -->|tenant| tenant
agent_version -->|agent| agent
agent_version -->|approvedBy| principal
approval_request -->|agent| principal
approval_request -->|approverRole| accessRole
approval_request -->|decidedBy| principal
approval_request -->|run| agentRun
tool_call -->|approval| approvalRequest
tool_call -->|run| agentRun
tool_call -->|tool| agentTool
tool_grant -->|agent| agent
tool_grant -->|grantedBy| principal
tool_grant -->|tool| agentToolConcepts
| Concept | Definition | Attributes | Written in |
|---|---|---|---|
agent | A program that decides for itself what to do, accountable to an owner and limited by its level of autonomy | 9 | commons/data/models/identity-agent/mod.ts:40 |
agentBudget | What an agent, under a delegation or in a run, may spend of a meter in a window; it cannot be exceeded | 9 | commons/data/models/identity-agent/mod.ts:165 |
agentRun | One task of an agent, in one version, perhaps under a delegation, perhaps started by another run | 7 | commons/data/models/identity-agent/mod.ts:114 |
agentTool | A tool that agents may be granted: what it is called, the permission a call needs, how risky it is | 8 | commons/data/models/identity-agent/mod.ts:81 |
agentVersion | A version of an agent: which model, which prompt and which tools, pinned, so that a bad one can be pulled from every run | 12 | commons/data/models/identity-agent/mod.ts:59 |
approvalRequest | A request for a person to approve one action of an agent, bound to a hash of its arguments, usable once | 14 | commons/data/models/identity-agent/mod.ts:184 |
toolCall | A call of a tool in a run: what was decided, and a hash of its arguments, never the arguments | 17 | commons/data/models/identity-agent/mod.ts:138 |
toolGrant | A tool that an agent has been granted, perhaps with a condition on its calls, never more than the agent's owner may do | 5 | commons/data/models/identity-agent/mod.ts:99 |
Code lists
| Code list | Definition | Codes |
|---|---|---|
approvalStatus | Where a request for approval is | 6 |
approverRule | Who may approve | 3 |
autonomyLevel | How much an agent may do without a person | 4 |
callDecision | What was decided about a tool call | 4 |
callStatus | Where a tool call is | 4 |
meter | What a budget counts | 4 |
runStatus | Where a run of an agent is | 6 |
toolApproval | Whether a tool grant needs approval | 2 |
versionStatus | Where a version of an agent is | 5 |
windowKind | How long a budget lasts | 5 |