Skip to content

Part VIII. Agent System Lifecycle

Up to this point, the book explained how to assemble the architecture, secure it, observe it, and roll changes out safely. But production discipline does not end at go-live.

Once an agent system lives longer than a single demo, a different class of questions appears:

  • which changes should count as release-significant;
  • how to respond to drift and findings;
  • how to preserve the lineage of trusted artifacts;
  • how to retire the system;
  • how to keep control over the whole estate, not only one agent.

This part answers exactly those questions. It reads the agent system no longer as an architectural diagram, but as a governed lifecycle.

Case thread in this part

The support-triage case now runs through the whole lifecycle arc: the duplicate-ticket fix becomes an ADLC change set, high-risk change packet, managed assurance finding, approved artifact bundle, retirement control for the old ticket writer, misalignment/control-eval scenario, detection-ready telemetry, and registry record with an owner. That lets the reader see that one incident should change not only code, but evidence, rollout, operations, and accountability.

Short path through this part

If you want a fast pass, read it this way:

  • Chapter 19: move from SDLC to ADLC as a working frame;
  • Chapter 20: decide which changes are release-significant;
  • Chapter 21: see how findings turn into response;
  • Chapter 22: lock down the lineage of trusted artifacts;
  • Chapter 23: close the lifecycle through replacement and retirement;
  • Chapters 24-27: extend the same contour into adversarial pressure, judgment, observability, and estate-wide accountability.

What This Part Solves

  • it shows the agent system as a governed lifecycle rather than a one-time launch;
  • it separates release judgment from response, lineage, closure, and estate accountability;
  • it gives the reader a language for change review, incidents, retirement, and sprawl;
  • it helps the reader view a production agent estate as a system with ownership, not as a pile of controls.

Role Map for This Part

Use this map to keep the late-book chapters from sounding like the same governance idea repeated under different names:

Lifecycle function Primary job Main artifact What it is not
Lifecycle frame Holds state transitions from design to retirement ADLC state model A new name for SDLC only
Change management Decides which changes require review and rollout gates Change packet General project management
Assurance Turns findings into containment, remediation, and ownership Finding and response record Observability or eval scoring
Provenance Preserves trusted artifact lineage and release identity Approved artifact bundle A generic evidence folder
Retirement Closes or replaces systems without losing accountability Retirement plan Deleting an old agent
Misalignment and insider risk Names adversarial or incentive-driven misuse paths Risk scenario and control plan A duplicate of prompt-injection guidance
Behavioral/control evals Produces release judgments about behavior and controls Eval gate and verifier contract Incident response
Observability Keeps the evidence substrate visible and queryable Trace and telemetry coverage record The owner of governance decisions
Inventory and registry Makes the estate answerable through owners and lifecycle state Registry record A loose spreadsheet of agents

Read the chapters as a chain: lifecycle defines the states, change management controls movement, evals judge readiness, provenance records what was trusted, observability preserves evidence, assurance responds when evidence turns into risk, retirement closes old paths, and registry keeps the whole estate accountable. If two chapters sound like "governance," distinguish them by the artifact they are responsible for leaving after review.

In This Part

What You Should Take Away

  • a more mature frame for release gates and change review;
  • a clear distinction between judgment, response, lineage, observability, and accountability;
  • a practical model for how an agent system changes, gets constrained, gets investigated, and eventually gets shut down over time.