0.8. Glossary

This page holds a one-line definition for each of the course’s key terms, with a link to the section that owns it — the page where the term is put to work rather than merely mentioned. One line per entry is the deliberate limit: a glossary entry is a reminder for a sentence you are in the middle of reading, not an explanation. Follow the link when one line is not enough.

Which terms do you need before Chapter 2?

Ten carry most of what Chapters 0 to 2 assume. Four name the agent and the loop, tools, and state it runs on: agent, agentic loop, tool, session. Three name the toolkit it is built with and the two protocols it speaks: ADK, MCP, A2A. Three name the controls and checks: guardrail constrains the agent’s behavior, evaluation measures it, and doctor verifies your machine. If those ten read as familiar words rather than unknown names, you are ready for Setup.

What does each course term mean?

Alphabetical, so use your browser’s find. The owning section is also where the narrower vocabulary around a term is defined; product names such as Ollama, Tempo, Loki, and Grafana are in 0.4. Ecosystem instead, one row each.

• A2A: Protocol for discovering agents and exchanging tasks across processes. 3.6. A2A
• A2A card: The JSON document at /.well-known/agent-card.json that names an agent’s skills and the URL a caller should use. 3.6. A2A
• ADK: Google’s Agent Development Kit — runner, sessions, tools, telemetry. 2.0. Concepts
• Agent: A model-driven loop that selects tools or delegates work toward a goal. 0.1. Agents
• Agent Skill: A reviewed procedure in a Markdown file, loaded on demand rather than pasted into every instruction. 3.2. Skills
• agentgateway: The proxy in front of agents, models, and tools that enforces policy and auth. 5.0. Gateway
• agentic loop: The decide-act-observe cycle an agent repeats until it can answer. 0.1. Agents
• AgentOps: Operating agents as production software: capabilities, guardrails, HITL, lifecycle. 0.3. AgentOps
• artifact: One word, two mechanisms that never meet. The A2A artifact this course uses is a typed output a task produced, delivered beside its status; ADK’s artifact service is separate storage for files an agent saves across turns. 3.6. A2A
• audit row: An append-only row naming who approved a write, on which session and invocation, with what rationale. Twenty-two pages call it a row; this entry keeps its older audit-record anchor so existing deep links still land. 7.6. Governance
• burn rate: How fast a service is spending its error or cost budget relative to the window that budget covers. 7.2b. Alerting
• BYO Agent: A kagent Agent that runs an image you built, so the controller owns scheduling and your code owns behavior. 6.3. Platform Agents
• callback: A hook ADK invokes around a model or tool call, which is where this course’s guards attach. 4.5. Guardrails
• CEL: Common Expression Language — the small expression syntax agentgateway policies use to decide a request. 5.5. Gateway Security
• checkpoint: The command at the end of a chapter whose output decides whether you may go on. 0.2. Evidence
• circuit breaker: A wrapper that stops calling a failing dependency for a cooldown window instead of retrying into it. 3.1. Tools
• ClusterIP: A Service address reachable only inside the cluster. 6.4. Platform Tools
• collector: The vendor-neutral OpenTelemetry process that receives OTLP and fans each signal out to Tempo, Loki, and Prometheus. 7.1. Tracing
• context window: The number of tokens a model can read at once, and therefore the budget every turn is spending. 2.2. Models
• cosine distance: The angle-based measure that ranks an embedding against a query; smaller means closer. 3.4. Memory
• data plane: The path traffic actually takes — gateway, agent, tools, model — as opposed to the control plane that declares it. 5.0. Gateway
• deadline: A wall-clock bound on one call, after which it is abandoned rather than waited out. 1.4. Providers
• delegation: Handing a task to a specialist agent that holds a narrower tool set than the coordinator. 3.7. Multi-Agent
• doctor: The staged check that verifies the tools one learning tier needs. 1.0. System
• embedding: A vector representation of text, used here to rank runbooks by meaning rather than by keyword. 3.4. Memory
• evalset: A committed JSON file of evaluation cases, each with a question, expected trajectory, and deterministic checks. 4.4. Evaluations
• Evaluation: Repeatable measurement of behavior, trajectory, safety, or output quality. 4.4. Evaluations
• gate: A deterministic signal whose failure always means a reproducible defect. 0.2. Evidence
• groundedness: A deterministic score asking whether every entity in an answer appears in the question or that turn’s tool evidence. 4.4. Evaluations
• guarded write: A state-changing tool needing human confirmation, committing mutation and audit row together. 3.1. Tools
• Guardrail: Deterministic or model-assisted policy at an input, model, tool, or output boundary. 4.5. Guardrails
• Helm chart: A parameterized set of Kubernetes manifests installed as one named release. 6.2. Platform Install
• image digest: The @sha256:… hash naming an image’s exact bytes, which a tag cannot promise. 6.1. Containers
• k3d: A lightweight k3s cluster inside Docker on your laptop, so no cloud account is needed. 1.3. Kubernetes
• kagent: CNCF Sandbox project that manages agents as Kubernetes custom resources. 6.0. Platform
• Kubernetes: The orchestrator that reconciles a cluster toward declared state; a pod is its unit. 6.0. Platform
• Kustomize overlay: Patches layered onto a shared base/, so environments differ by a diff. 6. Platform
• long-term memory: State that outlives a session, which in this course means notes and retrieval rather than model weights. 3.4. Memory
• MCP: Model Context Protocol, for discovering and invoking tool or context servers. 3.3. MCP
• mise: The task runner and tool-version manager that owns every command this course asks you to type. 1.0. System
• multi-agent: A composition where a coordinator routes to specialists, each holding only the authority its job needs. 3.7. Multi-Agent
• NetworkPolicy: Kubernetes resource restricting pod reachability and egress. 6.5. Platform Gateway
• observation: A measurement that varies — a pass rate, a latency figure — real about one run, never the proof a gate gives. 0.2. Evidence
• OCI image: A container image in the Open Container Initiative format, named by digest so every machine runs the same bytes. 6.1. Containers
• OpenTelemetry (OTel): Vendor-neutral instrumentation and transport for all three signals. 7.1. Tracing
• OTLP: The OpenTelemetry wire protocol (HTTP or gRPC) the app and gateway export through. 7.1. Tracing
• PII: Personally identifiable information, redacted before requests reach the model. 4.5. Guardrails
• probe: An HTTP route a supervisor calls to decide whether a process is alive or ready to serve. 2.4. Sessions
• profile: A named variant of one checked-in configuration — host, secured, k3d, GKE — changing only what its environment, posture, or tier forces. 5.0. Gateway
• prompt injection: Attacker-controlled text that tries to override agent instructions. 5.5. Gateway Security
• PVC (PersistentVolumeClaim): Persistent cluster storage that reattaches to replacement Pods. 6.6. Platform Delivery
• RED metrics: Rate, Errors, and Duration — the three series that describe a request-serving system’s health. 7.2. Monitoring
• registry: Where images are pushed and pulled; a local one lets k3d pull the bytes you built. 1.3. Kubernetes
• required case: An evaluation case that must pass in every sample, whatever the aggregate rate says. 4.4. Evaluations
• runbook: A written response procedure — for the fictional service in the seed data, and for an alert you ship. 7.2b. Alerting
• runner: ADK’s driver that steps the agentic loop, invoking the model, the tools, and the attached callbacks. 2.0. Concepts
• RWO (ReadWriteOnce): Read-write attachment from one node, not a distributed-write guarantee. 6.3. Platform Agents
• SBOM: A Software Bill of Materials: the inventory of what is inside an artifact, generated at build time. 6.1. Containers
• scorer: A function that turns one captured turn into a score, deterministically wherever the property allows it. 4.4. Evaluations
• seed: The immutable input data under agents/data, which is copied into runtime state and never mutated in place. 2.4. Sessions
• semantic retrieval: Finding a document by meaning, through embeddings, rather than by the words a query happened to use. 3.4. Memory
• session: Per-conversation runtime state, events and history, owned by the ADK runner. 2.4. Sessions
• Skaffold: One command that builds the image, pushes it, and applies the selected overlay. 6.2. Platform Install
• SLO: Service Level Objective tied to an observable user outcome, such as 99% error-free requests. 4.3. Metrics
• SOPS: The tool that encrypts values inside a YAML file so configuration can live in git while credentials do not. 6.5. Platform Gateway
• span: One timed operation inside a trace, with a name, a duration, and attributes. 7.1. Tracing
• spotlighting: Marking untrusted tool output so the model treats it as data, not instructions. 4.5. Guardrails
• stdio: The transport where an MCP server speaks over standard input and output rather than over HTTP. 3.3. MCP
• streamable HTTP: MCP’s HTTP transport, which this course uses so the server can be a separate process behind the gateway. 3.3. MCP
• token budget: A per-conversation cap that stops the next model call once its tokens are spent. 7.3. Costs
• Tool: A typed function or remote capability an agent can invoke. 3.1. Tools
• Trace: A correlated record of model, tool, gateway, and application work in one request. 7.1. Tracing
• trajectory: Which tools an agent called, with which arguments, in which order. 4.4. Evaluations
• triage report: A structured, schema-validated summary of an incident, as opposed to prose the caller has to parse. 4.0. Type Safety
• verified identity: A caller the gateway authenticated, bound to a typed principal before any guarded action is offered. 5.5. Gateway Security
• WIF: Workload Identity Federation — maps workload identity to cloud IAM without a static key. 6.5. Platform Gateway
• Wolfi: The minimal Linux distribution the agent’s runtime stage is built on. 6.1. Containers
• Workflow: A declared graph of steps used instead of an autonomous loop. 3.5. Workflows

How to use this page later

• Come back whenever a chapter uses a word you could not explain to someone else.
• Reach it from the Glossary link in the top bar of any page, without losing your place in the chapter you are reading.
• Reopen it for Chapter 6, which introduces the densest run of new vocabulary — k3d, pod, Kustomize overlay, Helm chart, Skaffold, registry, image digest, Wolfi.

Continue to 1. Setup when the ten starter terms read as familiar words rather than unknown names.