Chapter 16 placed a person above a small team of agents. This chapter adds a layer: agents that assign tasks to other agents, check what they return and route the result onwards, until the last output reaches a person as a file, a letter or a shift on a roster. Its problem is not that machines manage machines; it is that a chain of delegations can grow long enough for nobody to be able to say who answers for what it did to a person. The chapter concerns M1 Direction, which an orchestrator performs for its agents and sometimes for people; M2 Evaluation, which can quietly spread from agents to the people downstream; and M3 Consequence, which no orchestrator may hold at any depth of the chain. It ends with a testable answer to the open question Q7.
Vendors describe orchestration in terms that read like a supervisor's job description. Anthropic names the orchestrator-workers workflow, in which "a central LLM dynamically breaks down tasks, delegates them to worker LLMs, and synthesizes their results", and the evaluator-optimiser, in which one model generates and another evaluates in a loop (22, Anthropic, 2024). In management terms the first is M1 Direction of agents and the second M2 Evaluation of agents.
Anthropic's account of its own research system shows the pattern at production scale. A lead agent develops a strategy and spawns sub-agents that search in parallel; on the company's internal evaluation, a larger lead model with smaller sub-agents outperformed the larger model alone by 90.2 per cent (79, Anthropic, 2025). The account is equally frank that early versions spawned fifty sub-agents for simple queries and that "small changes to the lead agent can unpredictably change how subagents behave", the problems of a new supervisor: span of control, over-delegation and a team whose behaviour depends on how work is handed out.
Primary topics explored
- Agents managing agents
- Protocols and plumbing
- How multi-agent systems fail
- Many hands
- The moral crumple zone
- Where the chain ends
About AI Boss
AI Boss is Book 7 of The Transformed Architecture. It examines what happens when artificial intelligence takes the manager's chair: algorithmic management of direction, evaluation, consequence, development and care, the effects on people, the management of AI agents, Europe's legal answer, and how to design an accountable AI boss, for European enterprises and public authorities.
It builds on the conclusions of Architecture as Code, AI Governance, AI Platform and Security, AI Red Team, AI Threat and The Transformed Security Architect.
Executives, line managers, HR and people leaders, works councils and trade unions, public employers, architects and AI governance leads, data protection and legal officers, procurement and programme leaders, and risk, compliance and audit.
Gunnar Nordqvist
Conceived, directed and published the books of The Transformed Architecture.
Continue with the full chapter
Chapter 17: Orchestrators: AI Managing AI Agents and the Chain of Accountability in AI Boss gives the detailed explanations, practical examples and implementation patterns behind this summary.
This whitepaper belongs to “Managing Machines” and is one of 37 drawn from the manuscript.