0.7. Glossary

Which terms do you need before Chapter 2?

Ten terms carry most of what Chapters 0 to 2 assume. Each one links to its full entry below.

• Agent: a model that chooses tools in a loop until it can answer — here, an on-call assistant investigating an incident.
• agentic loop: decide, act, observe, repeat: the model asks for a tool, the runtime runs it, the result shapes the next step.
• Tool: a typed function the agent may call, such as get_service_status.
• session: one conversation's history and state, so the next turn knows what the last one said.
• ADK: Google's Agent Development Kit, the framework that runs the agent, its sessions, and its tools.
• MCP: the protocol that lets an agent call tools living in a separate server instead of in its own process.
• A2A: the protocol two agents use to find each other and exchange tasks across process boundaries.
• Guardrail: a check that runs in code around the model or a tool — for example, holding restart_service until a human approves it.
• Evaluation: a repeatable test of what the agent did, including which tools it called and in which order.
• doctor: the staged mise run doctor check that says whether your machine has the tools the next tier needs.

What does each course term mean?

Every term links to the section that introduces it. The list is alphabetical, so use your browser's find to jump straight to a word.

• A2A: Protocol for discovering agents and exchanging tasks across process boundaries. 3.6. A2A
• A2A card: JSON descriptor an A2A server publishes declaring its skills, endpoints, and streaming support. 3.6. A2A
• ADK: Google Agent Development Kit: the framework providing the agent runner, sessions, tools, and telemetry. 2.0. Concepts
• Agent: A model-driven loop that selects tools or delegates work toward a goal. 0.1. Agents
• Agent Skill: A small, discoverable instruction package rooted at SKILL.md, loaded only for a relevant task. 3.2. Skills
• agentgateway: The proxy placed in front of agents, models, and tools that enforces policy, auth, and observability. 5.0. Gateway
• agentic loop: The iterative decide-act-observe cycle an agent repeats — the model requests a tool, the runtime executes it, the result feeds the next step — until it can answer. 0.1. Agents
• AgentOps: Practices for operating agents in production: capabilities, guardrails, HITL, and lifecycle. 0.2. AgentOps
• audit record: An append-only row written in the same transaction as a guarded write, recording who approved which action against which target, so the change is attributable and never silently reversible. 3.1. Tools
• backoff and retry: Retrying transient read failures (cold start, restart, locked SQLite) instead of failing the turn. 3.1. Tools
• burn rate: How many times faster than planned a service is consuming its SLO error budget. 4.3. Metrics
• BYO Agent: A kagent "bring your own" Agent resource where the course ships its own A2A container image. 6.3. Platform Agents
• callback: An ADK hook around a model or tool boundary; this repository groups its policy hooks in one app plugin. 2.0. Concepts
• CEL: Common Expression Language, used here for boolean gateway authorization rules over each MCP request. 5.2. MCP Gateway
• checkpoint: The per-page verification exercise a learner completes to confirm the result before continuing. 0.0. Course
• ClusterIP: A Kubernetes Service address reachable only inside the cluster; use a port-forward for temporary host access. 6.4. Platform Tools
• circuit breaker: An opt-in reliability lever that opens after repeated read-tool failures so further calls fail fast, then lets one trial call test recovery. 3.1. Tools
• composition root: The one module where an application constructs its runtime and wires dependencies together. 3. Capabilities
• context window: The maximum tokens a model can attend to in one call — instruction, history, tool schemas, and tool results together; exceeding it fails the request or silently drops old content, depending on the serving path. 2.2. Models
• cosign: Tool used to verify an image signature and signed attestation. 6.1. Containers
• cosine distance: Embedding-similarity measure used by the optional semantic retrieval branch. 3.4. Memory
• cost attribution: Assigning model and token cost to a session or tool for budgeting. 7.3. Costs
• data plane: The proxy layer every request passes through, where routing, rate limits, and content policy are applied. 0.3. Ecosystem
• deadline: A bounded time limit wrapping read tools so a slow dependency costs a retry, not a hang. 3.1. Tools
• delegation: In-process transfer of control from a coordinator to a named sub-agent in the same process and session, cheaper and lower-latency than networked A2A but sharing one deployment, failure, and trust boundary. 3.6. A2A
• doctor: The staged prerequisite check (mise run doctor and its doctor:model, doctor:gateway, doctor:platform, doctor:gcp profiles) that verifies exactly the tools a learning tier needs before you start it. 1.0. System
• DSAR: Data-subject access request — a person asking what personal data you hold about them (or to erase it); the agent answers it per store, erasing memory while retaining the audit record. 7.6. Governance
• embedding: A vector representation of text that lets a search match meaning instead of exact words. 3.4. Memory
• efficiency metric: An informational evaluation value, such as tokens, model calls, tool calls, or duration, that never passes or fails a case. 4.4. Evaluations
• Evaluation: Repeatable measurement of behavior, trajectory, safety, or output quality — for example, checking that a question about an unknown incident calls get_incident with the right id and makes no unexpected write. 4.7. Evaluation Reference
• fallback model: An opt-in secondary model on the same provider, tried only after the primary returns 429 or 5xx before responding. 2.2. Models
• gate: A signal a machine evaluates deterministically, where a failure always means a defect a developer can reproduce and fix locally, so it can block a merge with no human in the loop. 4.3. Metrics
• Grafana: Dashboarding UI that visualizes the shipped AgentOps metric and log panels. 7.2. Monitoring
• Helm chart: A packaged, parameterized set of Kubernetes manifests installed as one named release; the course installs kagent this way, declared in infra/helmfile.yaml rather than typed at a prompt. 6.2. Platform Install
• image digest: The @sha256:… content hash that names an image's exact bytes; unlike a tag, it cannot be re-pushed to mean something else, which is what makes a build reproducible. 6.1. Containers
• groundedness: Whether an answer's claims are supported by retrieved evidence rather than invented. eval:ground deterministically checks recognized incident/severity and known service/runbook claims. 4.7. Evaluation Reference
• guarded write: A state-changing tool (restart_service, resolve_incident) that cannot run on model output alone: it requires human confirmation, validates its target, and writes its mutation and audit record in one transaction. 3.1. Tools
• Guardrail: Deterministic or model-assisted policy at an input, model, tool, or output boundary — for example, refusing a restart_service call until a human approves it. 4.5. Guardrails
• HITL: Human-in-the-loop: a human decision inserted before a sensitive action continues. 3.1. Tools
• k3d: Runs a lightweight k3s Kubernetes cluster inside Docker containers on your laptop, so the cluster chapter needs no cloud account. 1.3. Kubernetes
• kagent: CNCF Sandbox project that manages agents as Kubernetes custom resources. 6.0. Platform
• kill-switch: The AGENT_WRITES_DISABLED startup flag that makes every guarded write refuse before approval after the process restarts, while reads keep working. 4.5. Guardrails
• Kubernetes: The orchestrator that keeps declared workloads running: you describe the desired state in YAML and a controller reconciles the cluster toward it. A pod is its smallest deployable unit — one or more containers sharing a network identity and lifecycle. 6.0. Platform
• Kustomize overlay: A small folder of patches layered onto a shared base/ of manifests, so the local and gke environments differ by a diff rather than by two copies; kubectl kustomize <dir> prints the rendered result. 6. Platform
• Loki: Log store that receives OTLP logs and backs the dashboard's log panel. 7.2. Monitoring
• long-term memory: Explicit cross-session notes a stable user can save and recall, isolated per user and PII-redacted before the write, distinct from per-session conversation state. 3.4. Memory
• MCP: Model Context Protocol for discovering and invoking tool or context servers. 3.3. MCP
• mise: The task runner behind every mise run command in the course; it also pins each CLI to one exact version. 1.0. System
• MLflow: Self-hosted store for prompt versions, traces, and evaluations. 7.0. Reproducibility
• Multi-agent: A system where one agent delegates to several specialist agents under a coordinator. 3.7. Multi-Agent
• NetworkPolicy: Kubernetes resource that restricts pod reachability and egress to reduce attack surface. 6.5. Platform Gateway
• OCI image: The open container-image format Docker and Podman both produce: an immutable, content-addressable bundle of a filesystem and its start command. 1.2. Containers
• Ollama: The local model server that runs the course's model on your machine through an OpenAI-compatible endpoint. 0.4. Providers
• OpenTelemetry (OTel): Vendor-neutral instrumentation and transport layer for traces, metrics, and logs. 7.1. Tracing
• OTLP: The OpenTelemetry wire protocol (HTTP/gRPC) the app and gateway use to export telemetry. 7.1. Tracing
• PII: Personally identifiable information, redacted before requests reach the model. 4.5. Guardrails
• probe: A periodic kubelet check of startup, readiness, or liveness; it controls traffic or restart behavior, not application correctness. 6.3. Platform Agents
• progressive delivery: Releasing to progressively wider traffic groups with measured rollback criteria; this lab teaches only the promotion preflight and immutable rollback boundary. 6.7. Promotion and Rollback
• Prometheus: Time-series database that scrapes collector and gateway metrics. 7.2. Monitoring
• prompt injection: Attacker-controlled text that tries to override agent instructions. 5.5. Gateway Security
• prompt registry: MLflow-backed registry used to version and compare prompts in the host development/evaluation environment; production images use the committed instruction. 7.0. Reproducibility
• PVC (PersistentVolumeClaim): A request for persistent cluster storage that can outlive and reattach to replacement Pods. 6.6. Platform Delivery
• Qwen3: The course's default model, pulled as qwen3:4b-instruct, whose published weights are Apache-2.0 licensed. 0.4. Providers
• RAG: Retrieval that adds relevant external knowledge to the model context. 3.4. Memory
• RED metrics: Rate, Errors, Duration: bounded request metrics derived from spans under namespace agentops. 7.2. Monitoring
• registry: The server images are pushed to and pulled from; the course runs a local registry.localhost:5050 so k3d nodes pull the same bytes your build produced, with no internet round trip. 1.3. Kubernetes
• right to erasure: A person's request to delete their personal data; honored for long-term memory via forget_user_memory, while the append-only audit record is retained under a separate legal basis. 7.6. Governance
• RWO (ReadWriteOnce): A volume access mode allowing read-write attachment from one node; it is not a distributed-write database guarantee. 6.3. Platform Agents
• runner: The ADK object that turns one user message into one completed turn, driving the model, the tools, and the session. 2.0. Concepts
• SBOM: Software Bill of Materials listing an image's components for supply-chain auditing. 6.1. Containers
• scorer: A function that grades one recorded conversation; the five MLflow scorer means enforce floors inside a scheduled evidence run, not a merge gate. 4.3. Metrics
• seed (immutable seed): The committed agents/data fixture — incidents, service logs, runbooks, and Agent Skills — that the agent reads but never mutates. 2.1. First Agent
• semantic retrieval: Optional embedding-based search enabled only when it beats the keyword baseline. 3.4. Memory
• session: Per-conversation runtime state (events, history) owned by the ADK runner. 2.4. Sessions
• Skaffold: The build-push-deploy loop for Kubernetes: it builds the agent image, pushes it to the local registry, and applies the selected Kustomize overlay in one command. 6.2. Platform Install
• SLO: Service Level Objective tied to an observable user outcome (e.g. 99% error-free requests). 4.3. Metrics
• SOPS: Tool for keeping encrypted secrets in git and decrypting them at deploy time. 6.5. Platform Gateway
• span: A single timed unit of work in a trace, emitted by ADK and gateway instrumentation. 7.1. Tracing
• spanmetrics: OTel connector that turns spans into request-count and duration metrics. 7.1. Tracing
• spotlighting: Wrapping attacker-influenceable tool output in a marked prefix so the model treats it as data, not instructions. 4.5. Guardrails
• stdio: Default MCP transport where the agent launches the server as a child process. 3.3. MCP
• streamable HTTP: MCP HTTP transport option for a networked server. 3.3. MCP
• token: A chunk of text (~3–4 characters of English): the unit a model reads and writes, and the unit budgets and cost count. 2.2. Models
• token budget: A per-conversation cap that stops the next model call once a session's accumulated tokens are spent, bounding reasoning work rather than dollars. 7.3. Costs
• Tool: A typed function or remote capability an agent can invoke. 3.1. Tools
• tool annotation: An MCP hint that a tool is read-only, idempotent, or closed-world; clients display it but must not trust it as enforcement. 3.3. MCP
• 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 — what an agent evaluation scores besides the final text. 4.7. Evaluation Reference
• triage report: A schema-validated final answer produced by a second entry point for machine consumers such as tickets or dashboards, unlike the conversational agent's prose. 2.3. Instructions
• verified identity: A caller identity a trusted gateway proves by validating a JWT and forwarding it as a header, so the audit row names a real subject. 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, container-focused Linux distribution the agent's runtime stage is built on; it is a rolling repository, so its exact apk pins eventually need bumping. 6.1. Containers
• Workflow: A declared graph of steps used instead of one autonomous loop for deterministic control. 3.5. Workflows

How should you use this page later?

Come back here whenever a chapter uses a word you could not explain to someone else. Each entry gives you one line, then a link to the section that owns the term.

Keep the page open in a second tab while you read Chapters 1 to 3. That is when most of the unfamiliar names arrive, and a ten-second lookup beats guessing.

Reopen it for Chapter 6. It introduces the densest run of new vocabulary in the course — k3d, pod, Kustomize overlay, Helm chart, Skaffold, registry, image digest, Wolfi — and all of them are entries above.

You are done when:

• You can say what an agent, a tool, and a session are without opening another page.
• You followed at least one entry's link and landed on the section that introduces that term.
• This page is bookmarked, or open in a second tab, for the rest of the course.

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