Agent Engineering Glossary

You do not need to learn every internal module to read this site. Start with one distinction: a model can interpret and propose actions, while authorized systems execute them and retain receipts. These are terms used in the articles, not a new industry standard or runtime contract.

Execution and responsibilities

Term	Meaning in these notes	Further reading
Model	Interprets requests, proposes research actions, and drafts answers; its words alone cannot certify effects.	Six-layer map
Harness	Facilities connecting model decisions to work: tool visibility, constraints, state, and execution.	Model, Harness and Loop
Native Agent Loop	Receives model proposals, dispatches authorized tools, consumes real results, and continues or stops.	Responsibility split
Run	A task execution with its own identity, deadline, budget, and observed effects. Recovery retains this history.	Termination authority
Owner	The party responsible for particular facts, state, decisions, or effects—not necessarily a microservice.	Architecture principles
Entry / Research / Finalization	Stages for intake, advancing research, and concluding or delivering work—not three separate models.	Six-layer map
Main / Subagent / Child	Main worker and constrained task workers. Child results need attributed handoff before Main adopts them.	Subagent collaboration
Provider / Gateway	Model transport and protocol adaptation. An HTTP 200 response does not imply a completed task.	Provider admission
Tool Host / Tool Surface	The host validates and executes real tools; the surface is the set of tools actually visible to the model.	Tool Surface
MCP / Skill	MCP is an open integration protocol; a Skill gives on-demand instructions and resources. Connection alone does not prove adoption.	MCP lifecycle

Authority, evidence, and delivery

Term	Meaning in these notes	Further reading
Grant	Authorized scope, duration, or budget for a Run. Model proposals do not expand it.	Authorization
HIL / HIO	Human interaction and response adoption. Asking, receiving, adopting, and causing an effect are separate facts.	Durable human interaction
Publication	A versioned, identifiable, receipt-backed answer a user can retrieve, separate from tool execution success.	Answer publication (Chinese original)
Handoff	Transfer of a Child's output and identity to another task owner. Proposed, completed, and adopted states differ.	Subagent collaboration
Checkpoint / Recovery	Preserved execution state with original identity and already-observed effects; unknown is not automatically unexecuted.	Recovery authority
ResultRef / Artifact	Retrievable references to externalized results or files; the model's context is not the durable original.	Externalized results
Memory / RAG	Memory manages durable preferences or research effects; RAG retrieves and projects source content subject to freshness and revocation.	Memory · RAG
Evidence / Lineage	Traceable relationships among source facts, calculations, and claims. A citation alone does not prove a conclusion.	Evidence (Chinese original)
Eval / Dogfood	Task-specific quality checks, distinct from deterministic invariants and demonstrated user outcomes.	Testing and evaluation
Effect / unknown	Known persistent effects and unknown status remain distinct; a sent request without a receipt is not proof it never ran.	Principle three

External reference points

Owner, Publication, and Grant are terms in this engineering discussion; not all are standardized elsewhere. These resources help compare mechanisms, not claim implementation equivalence.

• Official MCP specification: tools, resources, and authorization concerns.
• Anthropic · Building Effective Agents: the trade-off between workflows and agents.
• Temporal · Workflow Execution: a reference for durable workflows; this project does not implement Temporal.
• ReAct paper: an early research framing for interleaving reasoning and actions.
• W3C PROV Overview: provenance concepts for entities and activities.

Start with the six-layer map · Runnable teaching project · Article catalog