AAuth Mission

Status: Exploratory Spec: draft-hardt-aauth-mission

Optional extension to the Protocol for multi-step agent workflows. Inspired by military Mission Command (Auftragstaktik) — communicate the objective, let the agent execute, evaluate each action against the approved intent.

What It Does

Mission Proposals

An agent submits a mission proposal to its Mission Authority (MA) — typically its own auth server. The MA evaluates the proposal (potentially deferring for human review or engaging in clarification chat) and returns an approved mission, identified by the SHA-256 hash of the approved text.

Mission-Scoped Authorization

As the agent works, it includes the mission identifier in each token request. The MA evaluates whether each resource access is consistent with the mission and can deny requests that fall outside scope.

Mission Control

The administrative interface to the MA. Because the MA is in the authorization path for every mission action, it holds a complete audit trail:


10:00  Agent → flights.example      scope: search        granted
10:02  Agent → flights.example      scope: fare.hold     granted
10:05  Agent → hotels.example       scope: search        granted
10:08  Agent → hotels.example       scope: reserve       granted
10:12  Agent → restaurants.example  scope: book          granted
10:15  Agent → payments.example     scope: charge        denied — outside mission scope

Users can suspend a mission (halting all access), resume it, or revoke it entirely.

The Analogy

OAuth’s model — pre-registered clients, static scopes, upfront consent — is Befehlstaktik (detailed command): specify every step in advance. This works when the environment is predictable.

AAuth missions are Auftragstaktik (mission command): approve the objective, let the agent execute, intervene when necessary. As Moltke observed, no plan of operations extends with certainty beyond first contact.

Primitives provided: scoped authorization contexts, centralized audit, mission lifecycle management