https://aibussin.com/categories/advanced-agents-from-first-principles/ Recent content in Advanced Agents From First Principles on AIBussin — AI applications, systems and books Hugo en-US Sun, 09 Aug 2026 16:09:00 +0100 https://aibussin.com/books/advanced-agents-from-first-principles/44-chapter/ Sun, 09 Aug 2026 16:09:00 +0100 https://aibussin.com/books/advanced-agents-from-first-principles/44-chapter/ <h1 id="build-a-production-ai-agent-from-first-principles-the-complete-reference-architecture">Build a Production AI Agent From First Principles: The Complete Reference Architecture</h1> <p>We have spent this series adding mechanisms only when a specific failure demanded them.</p> <p>We started with a model call.</p> <p>Then we added candidate generation, critique, planning, tool use, memory, search and verification.</p> <p>Then the system stopped looking like a clever prompt.</p> <p>It started looking like software.</p> <p>Then distributed systems problems arrived:</p> <ul> <li>duplicate work,</li> <li>retries,</li> <li>leases,</li> <li>fencing,</li> <li>backpressure,</li> <li>dependency failure,</li> <li>behavioral drift,</li> <li>release compatibility,</li> <li>replay,</li> <li>incident forensics,</li> <li>SLOs,</li> <li>authority,</li> <li>competence,</li> <li>capability acquisition,</li> <li>placement,</li> <li>handoff,</li> <li>stale state,</li> <li>stale intent,</li> <li>commitments,</li> <li>durable workflows,</li> <li>transaction recovery,</li> <li>trust boundaries,</li> <li>multi-agent coordination,</li> <li>and finally an explicit control plane.</li> </ul> <p>At this point the architecture is complete enough that adding another isolated mechanism would make the series worse rather than better.</p>