Book · First research edition
Designed to Be Debugged: why autonomous AI needs state capture, replay, and escalation before it can be trusted. An operational engineering discipline for agentic systems, translated from silicon debug practice.
The argument
The problem
Final-output monitoring misses the expensive part. The bills (incident, change, audit, scaling, and trust) arrive later, and they are paid only after an incident. Near misses are data, not noise; most systems throw them away.
The discipline
Chips ship with designed-in debug access: independent observation paths, meaningful state exposure, reproducible failure, and a threat model for the debug logic itself. The book translates that design-for-debug (DFX) discipline into agent architectures, before the tool boundary is fixed, not after the incident.
The architecture
Eight evidence surfaces turn an opaque sequence of model and tool interactions into an evidence-bearing operational run. A reference architecture: layers can be combined or renamed; what matters is preserving the questions each layer must answer.
The controls
One kill switch is not a control strategy. The Intervention Ladder names the response appropriate to the evidence, the consequence, and the point of commitment, and puts the checkpoint before commitment.
The seven action states
Each state is recorded independently, with its own event and timestamp. A failed or timed-out state can never be reported as success, and verification uses a source other than the agent's own claim.
Inside the book
Free · no email required
The book's back-matter checklist as a standalone reference: eight review gates, the seven action states, four outcome measures, and the go/no-go decision rule. Everything an engineering team needs to run the review this quarter.
The First Research Edition (v1.0.3, research cutoff July 10, 2026) is complete. I am currently entertaining conversations with technical publishers and AI infrastructure teams.