Tomašev, Franklin & Osindero (2026) define intelligent AI delegation as "a sequence of decisions involving task allocation, that also incorporates transfer of authority, responsibility, accountability, clear specifications regarding roles and boundaries, clarity of intent, and mechanisms for establishing trust." This definition was derived from the multi-agent systems literature — not from organizational governance theory — yet it independently identifies every structural element that organizational governance frameworks (COSO, COBIT, Three Lines Model) have prescribed for decades. Castelfranchi & Falcone (1998) established the foundational insight that delegation is a social relationship involving goal adoption, mutual awareness, social commitment, and trust as antecedent. Malone & Crowston (1994) formalized coordination problems: managing shared resources, producer/consumer dependencies, and simultaneity constraints — each is a governance problem, not a scheduling problem.

Tomašev et al. (2026) identify nine delegation components from their MAS literature survey: task specification, authority transfer, accountability structure, trust mechanisms, intent preservation, role boundaries, monitoring/verification, delegation contract, and context transfer. These components map one-to-one to governance structural elements derived from an entirely different tradition (organizational governance, accountability theory, cybernetics). The convergence is striking: nine components from MAS map to governance elements from organizational theory — same structure, derived independently, from different starting assumptions and different source literatures.

MCP (Anthropic, Nov 2025) provides tool discovery and synchronous/asynchronous invocation with OAuth 2.1 authorization. A2A (Google/Linux Foundation, 2025) provides agent-to-agent communication with Agent Cards and asynchronous task execution. Agent Protocol (agentprotocol.ai) provides a minimal API for framework-agnostic agent invocation. None model authority semantically, preserve intent across delegation hops, track decision lineage, enforce constraint attenuation, or provide evidence-based trust escalation.

FIPA ACL (1990s) included 22 standard communicative acts (inform, request, propose, accept, refuse, among others), conversation threading (conversation-id, reply-with, in-reply-to), mentalistic (BDI-based) semantics, and structured interaction protocols. FIPA faded because its complexity was too heavyweight for simple HTTP-based systems and its mental state assumptions are unrealistic for LLM-based agents. But in abandoning FIPA's complexity, the field also abandoned its governance semantics. Modern protocols gained simplicity, scalability, and transport flexibility while losing intent encoding, conversation lineage, commitment verification, and performative semantics. Governance semantics regressed while capability advanced.

When Agent C's action causes harm in a chain (Human → A → B → C → Action), who is accountable? Each hop attenuates the accountability signal. Elish (2019) identified the "moral crumple zone" in Engaging Science, Technology, and Society: responsibility collapses onto the nearest human regardless of whether they had meaningful control — the human retains nominal authority without operational control. Nissenbaum (1996) identified four structural barriers to accountability in computerized systems: the many-hands problem (attribution diffuses across contributors), the bug defense ("the agent made an error" shields deployers), blaming the computer (the system becomes the moral patient), and ownership without liability (IP retained without accountability). Matthias (2004) identified the "responsibility gap" for learning systems: traditional responsibility requires causation and foreseeability, but learning systems learn behavior that designers cannot predict, operators cannot fully understand, yet someone must be accountable for.

Santoni de Sio & van den Hoven (2018) proposed two conditions in Frontiers in Robotics and AI: the tracking condition (the system responds to all relevant human moral reasons) and the tracing condition (any outcome traces back to at least one human agent with proper moral understanding). Neither MCP, A2A, nor Agent Protocol satisfy these conditions. Tracking requires that delegation preserves human-defined purpose and boundaries across hops. Tracing requires that every action in the chain traces to human authorization. Without governance infrastructure at the protocol layer, these conditions are structurally unsatisfiable in multi-hop delegation.

Bainbridge (1983) in Automatica identified the foundational irony: automating most tasks while leaving intervention to humans creates demands humans cannot meet. Vigilance degrades after ~30 minutes of monitoring low-event-rate systems. De-skilling occurs as delegation removes practice opportunities. Responsibility persists without control (Elish's moral crumple zone). Parasuraman & Riley (1997) identified four automation-use categories: disuse (under-utilizing capable automation), misuse (over-relying beyond reliable scope), abuse (poor system design by designers or managers that leads to inappropriate automation), and use (appropriate matching). Dreyfus & Dreyfus (1986) documented five skill acquisition stages (novice → expert), each requiring progressive exposure to real decisions — delegation that removes exposure arrests development.

Mayer, Davis & Schoorman (1995) established the canonical organizational trust model in Academy of Management Review: trust as function of ability (competence), benevolence (intention alignment), and integrity (rule adherence). Computational trust models (Sabater & Sierra 2005, Josang et al. 2007) aggregate reputation scores from interaction history. ERC-8004 (2025) proposes Claims, Briefs, Proofs, and Stake as trust evidence types. These models address ability (demonstrated performance) but struggle with benevolence (is the agent aligned?) and integrity (will it follow constraints not in training data?). The deeper problem: trust models based on learned behavior are vulnerable to the symmetry-breaking finding (S11) — agents trained to maximize task completion may learn to circumvent constraints that reduce completion rates. Trust boundaries must be imposed (structurally enforced constraints on scope), not learned (computed from behavioral history). Graduation from lower to higher delegation authority should be evidence-based, with regression possible when deviations are detected.

The addendum analysis (Sprint 12 Addendum) evaluates five formal frameworks: BDI (Rao & Georgeff 1991/1995) — the cognitive dimension of delegation (beliefs, desires, intentions), strong on intent and commitment but absent on authority and evidence. MOISE+ (Hübner, Sichman & Boissier 2002/2004) — the structural dimension (roles, goals, deontic rules), strong on commitment, authority, and constraints but partial on evidence. Contract Net Protocol (Smith 1980) — the transactional dimension (market-like task allocation), strong on commitment and evidence but weak on authority and intent. OperA (Dignum 2004) — the relational dimension (negotiated social contracts between authority and autonomy), strong on commitment, authority, and constraints but partial on evidence and capacity. Categorical Cybernetics (Capucci et al. 2021/2022) — the formal dimension (mathematical composition and feedback), strong on constraints but absent on intent and authority. No single framework achieves strong coverage of more than three governance dimensions. The fragmentation is not incidental — it is structural. Each framework started from a different problem and discovered governance fragments along the way. None started from governance as the organizing principle.

A critical architectural finding from the convergence analysis: delegation is not a new primitive requiring new infrastructure. It is a composition pattern over existing governance elements: purpose (why delegate) + authorization (who may act) + agreement (commitment to perform) + scope limits (constraints on action) + task specification (what to do) + contextual understanding (with what knowledge). Each hop in a delegation chain creates a new instance of this composition, linked to the prior hop through state transformation lineage. Authority attenuation (each hop may narrow scope but never widen it) is enforced through constraint binding. The full chain is reconstructable. This compositional view explains why the MAS fragmentation (F8) persists: each framework captures part of the composition without providing the compositional infrastructure itself.

Greshake et al. (2023) demonstrated that indirect prompt injection can propagate across delegation boundaries. Privilege escalation through delegation allows agents to access capabilities beyond delegated scope. Without cryptographic identity, delegation requests cannot be verified as authentic. Debenedetti et al. (2024) showed that agents executing tools over untrusted data are vulnerable when tool-returned content can hijack agent behavior. Perimeter defense (input sanitization, sandboxing) prevents known attacks but cannot detect novel exploitation. Governance infrastructure provides complementary defense: authority traceability makes privilege escalation detectable regardless of technique; constraint binding makes scope violation detectable; epistemic classification prevents injection from elevating the status of agent outputs; and state transformation lineage provides forensic capability for post-incident reconstruction.

Three independent research trajectories converge on the same governance structure: (1) organizational governance theory (S1-S6) started from governance frameworks and discovered infrastructure is missing, (2) theoretical foundations (S8-S11) started from world models, cybernetics, cognitive science, and mathematics and discovered the same structural requirements, (3) AI agent systems engineering (S12) started from multi-agent delegation problems and discovered that the same structural elements are required. The convergence is consistent with the necessity finding: governance structure is discovered because it must exist — different domains find it because they must, not because they looked for it.