The Orchestrator

When Maya’s system produces a bad grouping, she does not start by rewriting the prompt. She works backward through the run, asking which conversations entered, which skills and memory were active, what the handoff carried, which tool ran, whether security review fired, and whether the eval set contained a case anything like this one. The answers map to layers:

• Evidence missing → context or retrieval.
• Evidence present, procedure wrong → skill.
• Wrong service called → tool description or capability boundary.
• Right recommendation, wrong branch touched → permission.
• Green CI on a test that never represented the customer → evaluation.
• Bad assumption passed downstream → handoff.

Inside a single platform, that diagnostic view is already shipping: Honeycomb’s Agent Timeline lays out the model calls, tool calls, and agent handoffs in a conversation in the order they happened, and New Relic and Arize map which agent handed off to which.³⁵ Distributed tracing can already join events across systems. Joining those events to the exact instructions, memory state, and authority active at each step still requires the operator to assemble and preserve the evidence. Chapter 16 describes the building blocks: the ability to diagnose multi-agent failures, to oversee interactions, and to see the decision chain with its instructions, data, and context at any point in time is the gate between orchestration as a productivity story and orchestration in mission-critical areas — power grids, hospital systems, banks.

The better analogy for all of this is a CI pipeline and a tighter OAuth scope, not a sterner paragraph. You teach a system that can invent its own steps by changing the path it is required to walk, and then by deciding whether a proposed lesson — a memory, a revised skill, a new test, a narrower permission — is safe to keep.

So the question to carry forward is not how to complete one task with an assistant attached. It is what environment would let a bounded system perform this task repeatedly: context, skills, tools, state, evaluation, authority. A prompt helps one run; the rest of the layers carry classes of runs. The instructions are part of that system, and they shape what the model does, but influence is not authority. In the hospital, the pharmacy’s limit holds because a validator checks every dose before anything is dispensed, whatever the agent made of its instructions. Those layers are what the next chapter’s delegation contracts and topologies have to describe.