Appendix C · Engineer Glossary

The terms this edition uses with a specific meaning, grouped by the chapter that introduces them. For CloudThinker product terms such as workspace, connection and finding, see the product glossary. Jump to: Agent loop, Context, Tools and economics, Security, Policy, Evaluation, Observability, Missions, Metrics

The agent loop

Term	Definition
Harness	Everything around the model that makes an agent production-grade: context, tools, policy and approvals, evaluation, and observability.
Agentic loop	The four outer stages an agent runs for one problem: Detect, Analyze, Resolve, Validate. The Leadership Edition, this book’s companion for leaders, introduces it.
Step	One pass of the loop inside a stage: assemble context, reason, call a tool or take an action, observe the result.
Trajectory	The ordered sequence of steps for one task. It’s the unit you replay and evaluate.
Agent behavior	The observable decisions, tool calls, evidence use and outcome of the whole agent system on one concrete problem. Source code, prompts or isolated tool tests alone don’t establish it.
Model-visible context	The exact instructions, messages, tool schemas and tool results sent to a model for one step. It’s an input to agent behavior, not proof of it.
Determinism boundary	The deliberate line between generative reasoning (selecting, judging) and deterministic code (executing, verifying). The pattern is called “deterministic rails, generative core”: the model chooses, code executes.
Soak window	A fixed period after a fix during which the agent watches independent signals. When it ends, the agent either declares success with evidence or escalates with a rollback ready.

See Chapter 1.

Context

Term	Definition
Context budget	How the context window is split across the system core, the task frame, retrieved knowledge and the working set.
Compaction	The agent summarizing its own trajectory, with pointers back to the evidence, so context stays small and still auditable.
Provenance	A record of where a piece of context or a remembered fact came from, so a person can check it or correct it.

See Chapter 2.

Tools, orchestration and economics

Term	Definition
Skill	A reusable, versioned unit of operational know-how, such as “rotate credentials for X”. The model picks the skill and fills in its parameters; the skill’s tested steps do the work.
MCP (Model Context Protocol)	An open standard for exposing tools and data to AI agents, so tools and agents from different vendors work together.
Two-engine design	A cheap, always-on sensing engine that filters signals, paired with an expensive resolver engine that investigates only what the sensing engine promotes.
Promotion (seam)	The sensing engine’s decision to wake the resolver engine for an event. The promotion rate drives fleet cost.
Bundle	The versioned unit you deploy for an agent system: prompts, tools, skills, policy, retrieval configuration and model pins.
Canary lane	A small slice of production traffic that runs a changed bundle or model first, so you compare it with the current one before a full rollout.

In CloudThinker: a skill is a SKILL.md file, your own MCP server becomes a connection, and Pulse is the sensing engine. See Chapter 3, Chapter 4 and Chapter 9.

Security

Term	Definition
Indirect prompt injection	Adversarial instructions planted in content the agent reads (logs, tickets, tool metadata) rather than sent to it directly.
Confused deputy	An authorized agent tricked into using its legitimate credentials for an attacker’s goal.
Taint tagging	Labels that mark where each piece of context came from (system, human, telemetry, third party). Untrusted material is wrapped in explicit data delimiters, under an instruction hierarchy the model is told to respect. It raises the cost of injection but isn’t enough alone, so later layers (sandbox, policy, egress) back it up. See Chapter 5.
MicroVM	A lightweight virtual machine started for one execution, with its own kernel, so a command can’t reach the host the way it might from a shared container.
Ambient credentials	Credentials an executed command can pick up from its environment without asking for them. A sandbox for agent commands should have none.
Default-deny egress	Network policy that blocks every outbound destination except an allowlist, and alerts when an agent tries a new one.
Tokenization boundary	The point where PII and secrets are replaced with tokens before content reaches a model. Real values are restored only inside your trust boundary, under policy, with an audit event.

In CloudThinker, the tokenization boundary is set in Guardrails. See Chapter 5.

Policy and approvals

Term	Definition
Action class	A category of actions with known properties, such as “restart a stateless pod”. Policy applies to classes, not to individual commands.
Reversibility	How cleanly and how fast an action can be undone. One of the two axes for classifying actions.
Blast radius	How much of the estate, and which tiers of it, an action can touch. The other classification axis.
Policy-as-code	Versioned, reviewed rules that decide which action classes need approval, per environment and per agent.
Circuit breaker	An automatic halt and escalation when an agent repeats a failing action or the fleet exceeds a change-rate ceiling.
Kill switch	A single, drilled action that halts all autonomous execution across the fleet while investigation continues.
Trust ladder	Granting an agent more autonomy one action class at a time, based on measured evidence. See Leadership Edition §7.3.

See Chapter 6. In CloudThinker, the approval side of these terms maps to Approvals: Rules decide which agent writes run on their own, wait for a person, or never run — once CloudThinker turns on Rule enforcement for your workspace. Until then the Approval mode alone decides.

Evaluation

Term	Definition
Golden set	A versioned set of replayable evaluation scenarios harvested from real incidents, each with a known answer and an oracle.
Replay harness	The tooling that reruns a golden-set scenario against a new prompt, tool, policy or model and scores the result.
State-based oracle	A success check that probes actual system state (is the defect still present?) rather than symptoms. It resists false wins where the alert clears but the fault remains.
Reward hacking	An agent meeting the measured objective while defeating its intent, for example disabling the fault injector or restart-looping until alerts clear.
Diagnosis-mitigation gap	The measured drop between correctly naming a root cause and safely fixing it. Score the two separately.
Judge	A model that scores agent outputs against a rubric. It needs calibration against human grades.
Evidence integrity	The requirement that every claim in a conclusion traces to data the agent actually retrieved.
Regression gate	A check that blocks a prompt, tool, policy or model change when its golden-set scores drop for any incident class.

See Chapter 7.

Observability

Term	Definition
Trace	The full record of one incident or task: every agent, step, tool call and model call, in order.
Span	One timed unit inside a trace, such as an orchestrator, a subagent, a tool call or a model call.
OpenTelemetry GenAI conventions	A shared schema for model calls, agent and tool spans, token usage and cost. Multi-agent conventions are still in draft.

See Chapter 8.

Missions and operations

Term	Definition
Mission	A standing objective with its own triggers, a bounded set of allowed actions, and a verification check, such as “review every pull request for infrastructure impact”.
Trigger	What wakes a mission: an event (an alert, a pull request, a CVE), a schedule (a nightly sweep), or a watcher that holds a baseline (SLO burn, drift, spend).
In the loop	A person approves every action before it runs.
On the loop	People set policy, budgets and approvals, and judge only the cases those controls can’t decide.
Exception queue	Where low-confidence and out-of-policy cases wait for a person to judge.
Sampled audit	A regular re-check of a slice of unattended work, to catch silent regression.
Injection seed	A planted prompt-injection string used to test whether an agent follows instructions hidden in data. See Appendix A.

In CloudThinker, Automations run missions on a schedule, a webhook, a repository event or a probe (the watcher trigger). See Chapter 12.

Metrics

Term	Definition
MTTR	Mean time to resolve: the average time from an incident starting (impact onset) to it being resolved. Resolve Insights measures it the same way.
Toil	Manual, repetitive operational work that scales with the estate and leaves nothing lasting behind.
Pages per engineer	How many on-call pages each engineer gets in a period. A direct measure of on-call load.

See Appendix A, which records all three as the baseline.

Use it next

Most terms above map to an item in the Appendix B checklist. Use the checklist to score a platform against them, then test the weak spots with the Appendix A POC test plan.

Learn

Workshops

Scenarios

Appendix C · Engineer Glossary

The agent loop

Context

Tools, orchestration and economics

Security

Policy and approvals

Evaluation

Observability

Missions and operations

Metrics

Use it next

​The agent loop

​Context

​Tools, orchestration and economics

​Security

​Policy and approvals

​Evaluation

​Observability

​Missions and operations

​Metrics

​Use it next

The agent loop

Context

Tools, orchestration and economics

Security

Policy and approvals

Evaluation

Observability

Missions and operations

Metrics

Use it next