Volume 1 Operational Glossary

Concise definitions anchored to the chapter where each concept is developed. When a term takes on distinct nuances across different volumes of the series, it is noted. This glossary does not substitute for the chapters; it resolves ambiguities mid-reading and establishes a shared baseline vocabulary for your engineering team.

Term	Operational Definition	Chapter
Agent	A software system that governs the control flow of its own steps based on task state, not merely the content of its textual output. Distinguished from a workflow by who owns control flow decisions.	1, 4
Workflow (Fixed Chain)	A predefined sequence of execution steps; the language model synthesizes content but does not determine the path. Budgeted, tested, and governed differently than an autonomous agent.	1
Autonomy Spectrum	A formal scale of levels (0 to 4) classifying a system by how much decision authority it exercises autonomously. Selected based on task structure, never organizational ambition.	1
Harness	The suite of deterministic runtime controls surrounding the model: iteration budgets, authorized tools, output schema validators, and loop breakers. The critical layer that allows an experimental agent to survive in production.	4
ReAct	A reasoning-action-observation loop alternating cognitive deliberation with tool execution until an objective is met. Foundational baseline for the family of agent reasoning patterns.	4
Context Window	The finite token capacity available in a single inference call. It is not persistent memory: it purges completely upon turn completion unless backed by an external persistence layer.	4, 10
Agent Memory	Persistent state that outlives an individual inference call, retrieved via explicit mechanisms (key-value stores, vector indexes, episodic summaries). Governed by an observable lifecycle.	10
Tool	A capability exposed to an agent with an explicit name, description, and input schema. The description guides invocation; the host server strictly enforces authorization and execution boundaries.	4, 9
Skill	A discrete unit of procedural knowledge with progressive disclosure: metadata is always accessible, while detailed operational instructions load only when task context demands it.	7
Progressive Disclosure	A context optimization pattern ensuring only essential operational detail is injected into the prompt at any given step, minimizing latency, token cost, and model degradation.	7
Synthetic Enterprise	The organizational pattern of decomposing complex workflows into specialized agent roles with strict contracts, counteracting the entropy of monolithic “god agents.”	5
Master Prompt	A single, authoritative prompt template reused across multiple agents, applying DRY to prompt engineering. Introduces architectural coupling that must be actively governed.	5
Bounded Context	A domain boundary within which an architectural model and ubiquitous language maintain unambiguous meaning. Serves as an immune system against semantic drift.	6
Domain Invariant	A business rule that must hold true across every state transition in an aggregate. An agent can violate it without detection unless explicit deterministic validators enforce it.	6
Orchestration Pattern	A canonical topology for coordinating multiple agents: pipeline, fan-out, council, hierarchical delegator, or self-reflection loop. Each mandates its own graceful degradation policy.	8
Graceful Degradation	The disciplined operational posture of serving reduced but explicitly declared quality or coverage rather than suffering hard failure. Degrading silently is lying with infrastructure.	3, 8
Model Context Protocol (MCP)	An open standard converting $M \times N$ custom integrations into an $M + N$ architecture over JSON-RPC, exposing standardized tools, resources, and prompts.	9
AI Gateway	A core platform infrastructure component centralizing model routing, API key management, quota allocations, rate limiting, semantic caching, and audit logging.	3
Instruction Corpus	The version-controlled repository of Markdown specification files governing agent behavior. Managed with the same rigor as source code: reviewed, tested, and audited.	10
Dependency Rule (in Context)	The architectural principle dictating that instructions must depend inward toward stable abstractions rather than volatile details, mirroring Clean Architecture.	10
Deterministic Verifier	A non-probabilistic gate (unit test, schema validator, linter) verifying agent output. The second line of defense complementing prompt context.	6
Eval	An automated, reproducible benchmark measuring probabilistic behavior against verifiable rubrics; replaces binary unit tests where outcomes are non-deterministic.	2, 4