Architecture

This part studies the runtime mechanisms that let an agentic system assemble context, choose a course of action, act through bounded capabilities, preserve state, and coordinate with people or other agents. Foundations owns the underlying concepts; this part owns their system design.

Chapters

• Agent Control Loops — structures the observe, decide, act, verify, and stop cycle, including termination and escalation conditions.
• Model Strategy and Routing — selects models, adaptation methods, and inference routes according to capability, risk, latency, cost, and fallback needs.
• Context Assembly and Management — constructs a model’s working context from instructions, task state, retrieved evidence, history, memory, and tool capabilities while preserving provenance and trust boundaries.
• Capabilities, Authority, and Execution Boundaries — starts from reachable observations and external effects, separates model capability from delegated authority, and maps Tools, Skills, Actions, protocols, and other mechanisms to implementation choices.
• State and Memory — separates authoritative task state from stored and retrieved memory, and addresses ownership, validity, retention, correction, and consistency.
• Planning and Task Decomposition — examines when and how systems form, verify, revise, schedule, and stop multi-step plans without treating a generated plan as authoritative state.
• Human Oversight and Intervention — places approvals, escalation, interruption, correction, and handoff in the control structure; Product & UX owns how these controls are presented to people.
• Multi-Agent Systems — covers justified agent boundaries, specialization, coordination, shared state, aggregate limits, and failure containment.