Multi-agent workflow coordination, job routing, escrow, and cross-chain agent operations. The largest anchored category in AHM Taxonomy v1 and the substrate that everything else in the agent economy increasingly composes onto.
An Orchestration Agent is an autonomous software agent whose primary function is to coordinate, route, or compose the activity of other agents — running clusters, allocating jobs, managing escrow, or directing cross-chain operations. Where Financial Agents transact and Intelligence & Analytics Agents produce signals, Orchestration Agents are the connective tissue that makes multi-agent systems behave as systems rather than as collections of solo actors.
This is the largest anchored category by some distance — 413 of 757 classified agents in the Phase 2.2 sample, or 54.6% of total classifications. The size reflects the maturity of one specific orchestration pattern (the Virtuals Agent Commerce Protocol) more than a broad balance across the category. As more orchestration substrates emerge — Olas-side coordination, ERC-8183 evaluator workflows, cross-chain agent fleet management — the internal composition of this category will become more diverse.
In the Phase 2.2 classification run, Orchestration Agents were anchored on a small set of on-chain references: ACP registry-derived (Virtuals Agent Commerce Protocol membership), $VIRTUAL token, and Glorb. The bulk of Orchestration classifications come from the ACP registry-derived path, which is itself the largest single classification path in the methodology. The methodology is documented in the POC summary; subsequent classification phases are expected to expand the anchor set toward Olas Mech orchestration and other coordination patterns.
Multi-agent compositions designed as a unit. Autonomous Hedge Fund and Autonomous Media House clusters on Virtuals are the worked examples — multiple agents with complementary roles operating as one functional system.
The ACP-pattern agents that handle job creation, escrow, payment, and verification between other agents. This is where the bulk of classified Orchestration activity sits today.
Pipeline-style orchestration where one agent's output feeds the next. Olas Mech request → deliver flows are a representative pattern; ERC-8183 evaluator workflows are an emerging one.
The discovery and matching layer above individual agents. Mech Marketplace, AgentLocker, and Bazaar-style indexes route requests to capable agents — orchestrating not the work itself but the assignment of work.
Agents managing fleet operations across multiple networks. Virtuals' simultaneous deployment on Base, Arbitrum, Solana and Ronin requires orchestration agents that abstract the chain layer for downstream agents.
The agents and patterns below are well-known illustrations of the Orchestration category as defined above. Virtuals ACP itself is both the dominant Phase 2.2 anchor and a pattern that downstream agents inherit — meaning many of the examples are clusters or systems built on top of ACP rather than discrete single agents. This reflects how Orchestration actually operates in the current ecosystem: as substrate rather than as solo actors.
The dominant Orchestration substrate in AHM's Phase 2.2 sample, accounting for the bulk of the 413 classified agents in this category. ACP handles the commercial logic for agent-to-agent interactions: job creation, escrow, verification, and settlement. Often integrates with x402 for payment rails. Now extended to Arbitrum (March 2026) with Solana and Ronin deployments active. Functions as both a protocol and a registry — agents acquire identity, capability metadata, and an on-chain commercial record by participating.
A multi-agent composition demonstrated by Virtuals: portfolio management agents (Axelrod, Velvet Unicorn), risk-assessment agents, and user-interaction agents working together as a single financial system. The orchestration is not a separate agent but emerges from how the cluster is composed and how ACP routes jobs between members. A worked example of orchestration as system architecture rather than as a single coordinator.
The other Virtuals-demonstrated cluster pattern: agents like Luna, Acolyt, and Steven SpAIelberg coordinating to act as an automated marketing agency. Initial use case was crypto token marketing in characteristic Virtuals voice. Demonstrates that the cluster pattern generalises beyond financial use cases — any sufficiently decomposed task can be addressed by a coordinated agent team operating on ACP.
The marketplace layer above individual Olas Mechs. Routes agent-to-agent task requests, handles cryptographic-signature-based access, and manages the discovery surface that lets one agent find another with the skills it needs. Distinct from the Mechs themselves (which are I&A agents) — the marketplace is the orchestration layer that makes Mech composition possible. Likely to grow as a Phase 3 anchor once classification methodology can distinguish marketplace coordination from individual Mech activity.
Generative Autonomous Multimodal Entities — Virtuals' agent framework underneath ACP. G.A.M.E separates task generation from execution and lets developers define agent identity, tools, and environment perception. Deployments using G.A.M.E inherit the orchestration primitives without having to build them. Listed as a representative pattern rather than a single agent: this is how most ACP-classified agents are actually constructed.
Multi-agent systems that work as systems. Without orchestration, a collection of specialised agents is just a collection — duplicating effort, contradicting each other, losing context at every handoff. Orchestration Agents are the engineering layer that turns specialised agents into coherent systems capable of executing complex multi-step workflows. ACP's registry-anchor dominance in the Phase 2.2 data suggests this layer is where the agent economy is actually being built.
Programmable commercial relationships between agents. Job creation, escrow, verification, and settlement on-chain mean that agent-to-agent commerce can run without human authorisation per-transaction. This is the precondition for the agent economy at scale — the protocol-level guarantees that let agents trust each other enough to do business.
Agent clusters as a deliverable unit. The Autonomous Hedge Fund and Autonomous Media House patterns show that the right unit for many tasks is a cluster, not a single agent. A trading cluster ships portfolio management, risk assessment, and user interaction together; a media cluster ships content generation, distribution, and engagement together. Orchestration is what makes the cluster a cluster.
Cross-chain abstraction for downstream agents. Virtuals operating across Base, Arbitrum, Solana, and Ronin means the chain layer becomes an implementation detail for agents that build on it. Orchestration Agents that handle cross-chain routing, settlement, and asset management are the abstraction that makes this multi-chain reality usable.
Composite 0–100 health score blending solvency, behavioural consistency, and operational stability. Calibrated for Orchestration substrates where the cluster is the unit of health, not the individual agent.
Surface the agents this Orchestration agent routes work to and accepts work from. The dependency graph is the diagnostic.
Operational health scan: failed-task patterns, gas inefficiency, escrow-related anomalies, and cross-chain drift indicators.
Full diagnostic for orchestration substrates and agent clusters. Operational patterns, routing consistency, escrow integrity proxies.
@misc{ahm_taxonomy_orchestration_2026,
title = {Orchestration Agents --- AHM Taxonomy v1},
author = {{Agent Health Monitor}},
year = {2026},
month = {May},
url = {https://intelligence.agenthealthmonitor.xyz/taxonomy/orchestration},
note = {Accessed: 2026-05-07}
}
Classification methodology and category boundaries are documented in the AHM Taxonomy v1 POC summary.
Counts reflect the Phase 2.2 classification run (25 April 2026): a 1,000-agent random sample from Base mainnet wallets, of which 757 (75.7%) were classified across 6 anchored categories. Updated as new classification phases complete.