About this eBook

Single-platform agents hit a wall the moment a process crosses departments. The answer is not one bigger agent — it is several small ones with clear contracts between them.

This book builds one complete system: an inbound enquiry arrives by phone, a voice agent qualifies it, an orchestrator writes to shared state, specialist workflows enrich and route it, a batch job reconciles overnight, and every step is traceable to a single correlation ID.

It also covers what nobody enjoys: what happens when step four succeeds and step five fails, how to make handoffs idempotent, how to test a system whose parts live on four vendors, and how to price the whole thing per transaction.

What you will be able to do

• Choose the right platform per step instead of forcing everything into one
• Design agent-to-agent contracts with explicit schemas and versioning
• Run shared state and correlation IDs across four systems
• Make every handoff idempotent so retries do not double-charge anyone
• Trace one transaction end to end across vendors and read the timeline
• Add approval gates, kill switches and blast-radius limits before you need them
• Model cost per transaction and find the step that is quietly expensive
• Evaluate the whole system, not just individual prompts

Table of contents

1. When one agent is not enough — decomposition and the cost of a hop
2. A platform decision matrix: latency, control, reach, batch, price
3. Contracts first — payload schemas, versioning and backward compatibility
4. Shared state: the store, the correlation ID and the event log
5. The orchestration layer and the router agent
6. Voice front door — ElevenLabs intake handing off cleanly
7. Private data work in self-hosted n8n
8. The long tail of SaaS through Zapier
9. Batch and reconciliation jobs in Make
10. Idempotency, retries, compensation and rollback across vendors
11. Distributed tracing and reading a cross-vendor timeline
12. Guardrails: approvals, rate limits, kill switches, blast radius
13. Evaluating a system — end-to-end suites, canaries and shadow runs
14. Cost per transaction and where the money actually goes
15. Runbooks, on-call and handing the system to a team

Who it is for

• Readers of the first four books ready to combine them
• Technical leads designing automation that spans several departments
• Consultants architecting systems they will have to support for years

Questions