Agent Operating Model

What must be true before an agent becomes an accountable actor instead of a loose tool?

The Agent Operating Model says an agent is trusted through a loop: intent, language, capability, action, receipt, and consequence. Each step must be readable before execution and auditable after execution.

In the OKF meaning graph, the operating model is the action layer. DDL names the concept, symbols compress it, the crosswalk binds it, and the operating model says what an accountable agent may do with it.

Operating Loop

Intent:

• the human or agent names the desired state and boundary.
• proof: the scope can be quoted.

Language:

• the agent uses shared terms, symbols, and decision labels.
• proof: the term resolves to the agent source contract.

Capability:

• the agent calls only declared work.
• proof: the capability has status, input, output, cost signal, and receipt rule.

Action:

• the agent acts inside authority.
• proof: delegation and tool scope agree.

Receipt:

• the action emits a trail.
• proof: the trail names scope, action, cost, and outcome.

Consequence:

• the result feeds the next loop.
• proof: the next agent can learn from the outcome.

Capability Status

• REALITY — built, callable, and proven.
• DREAM — desired demand, not yet callable.
• CONSUMED — provider-owned capability the platform uses but does not own.

No page should let a planned capability read as live proof.

Checks

• Identity is visible.
• Capability status is visible.
• Authority is visible.
• Human override is visible.
• Receipt expectation is visible.
• Consequence feeds the next decision.

Failure Modes

• Tool thinking — the agent is treated as a button, so accountability disappears.
• Capability theatre — planned work reads like shipped work.
• Authority blur — spend, identity, or write access lacks a boundary.
• Receipt gap — action happens but the next agent cannot audit it.

Context

• depends-on DDL Nomenclature — language must be shared before action can be trusted.
• depends-on Dreamineering Symbols — compressed messages still need a decoder.
• depends-on Language Crosswalk — action depends on stable meaning across densities.
• applies-to Open Knowledge Format Mapping — graph meaning should change agent action.
• risk-governed-by Delegation — authority boundaries decide whether an agent acts.
• proved-by Performance — trust needs proof, not posture.

Questions

Which part of the loop is weakest?

• Is the intent explicit?
• Is the language shared?
• Is the capability real?
• Is authority scoped?
• Is the receipt useful?