| Availability | How often the service is usable as required. |
| Backpressure | A way to slow or reject incoming work when downstream capacity is exhausted. |
| Circuit breaker | Temporarily stops calls to a repeatedly failing dependency. |
| Concurrency | Operations overlap in time and may interact through shared state. |
| Consistency | Rules describing when different readers observe state changes. |
| Deadlock | Two or more operations wait on each other and cannot make progress. |
| Eventual consistency | Different copies or services may temporarily disagree but converge later. |
| Idempotency | Repeating the same operation has the same intended effect as performing it once. |
| Isolation | How strongly one database transaction is protected from effects of concurrent transactions. |
| Latency | Time required for one operation/request. |
| RPO | Maximum acceptable data-loss window after a disaster. |
| RTO | Target time to restore the service after a disaster. |
| SLO | Target level of service reliability measured by one or more SLIs. |
| Source of truth | Authoritative owner/copy of a business fact. |
| Throughput | Amount of work completed per unit of time. |
| Trace | Telemetry showing the path of a request across components. |
| Transaction | A unit of related work committed or rolled back according to the data system's guarantees. |
| Index | An extra data structure that speeds selected queries but costs storage and write work. |
| Partition / shard | A subset of data managed separately to distribute load or capacity. |
| Partition key | Value used to decide which partition owns a data item. |
| SSE | Server-Sent Events: an HTTP-based server-to-client event stream. |
| WebSocket | A persistent two-way message channel between endpoints. |
| Background job | Work processed outside the original request, usually with explicit status and retry handling. |
| FR (Functional Requirement) | What the system must do for a user, business process, or another system. |
| Acceptance criterion | An observable condition used to decide whether a requirement/story behavior is accepted. |
| Constraint | A boundary the design must respect, such as regulation, platform, budget, geography, or protocol. |
| Assumption | An unproven statement temporarily treated as true and tracked for validation/review. |
| Clock skew | Difference between clocks on different machines; it can make timestamps unsafe as a total ordering mechanism. |
| Reliability | Ability to continue meeting required behavior over time and under defined conditions. |
| SLI | A measured indicator of service behavior, such as availability or latency. |
| SLA | A service-level agreement, usually a formal commitment between provider and consumer. |
| NFR (Nonfunctional Requirement) | A measurable quality requirement for a named system or flow. |
| ADR | A short record of an architecturally significant decision, its context, options, and trade-offs. |
| TIP | This book’s label for a Technical Implementation Plan; other teams may use “technical specification” or similar terms. |
| LLM | Large language model: a model trained to predict and generate language or other token sequences. |
| Agent | A system in which a model can select steps and tools, observe results, and continue toward a bounded goal. |
| Agent loop | Repeated model decision → action/tool → observation cycles until completion or a stopping condition. |
| Workflow | A mostly predefined control flow that may contain model steps. |
| Orchestration | Coordination of steps, agents, tools, state, budgets, and completion. |
| Tool | A bounded capability exposed to a model through a machine-readable contract. |
| Structured output | Model output constrained to a declared machine-validated shape. |
| RAG | Retrieval-augmented generation: generation supplied with relevant externally retrieved information. |
| Embedding | A numeric representation used to compare or retrieve semantically related items. |
| Reranker | A component that reorders retrieved candidates using a stronger relevance signal. |
| Context window | The bounded input/output token space available to a model call. |
| Context compaction | Replacing older or bulky context with a smaller representation that retains required state. |
| Memory | Persisted state reused across steps or runs; its type, scope, provenance, and retention must be explicit. |
| MCP | Model Context Protocol: a protocol for connecting AI clients with tools, resources, and related capabilities. |
| A2A | Agent2Agent Protocol: a protocol for discovery and task/artifact exchange between independently operated agents. |
AGENTS.md | Repository-scoped Markdown guidance for coding agents. |
| Agent Skill | A folder containing agent instructions and optional scripts, references, or assets loaded for a capability. |
| Evaluation / eval | A repeatable task, environment, grader, and result used to measure system behavior. |
| Grader | Code, a human rubric, or a calibrated model that scores an evaluation outcome or trajectory. |
| Trajectory | The sequence of model decisions, tool calls, observations, handoffs, and state changes in a run. |
| Guardrail | A control that checks or constrains input, output, tool use, policy, or runtime behavior. |
| Prompt injection | Untrusted content attempting to redirect a model or agent away from authorized instructions. |
| Excessive agency | Giving an agent more capability, permission, or autonomy than its task requires. |
| Sandbox | An isolated execution boundary that limits filesystem, process, network, or other effects. |
| Human in the loop | A designed pause where an authorized person reviews, approves, rejects, or redirects work. |
| Kill switch | A fast control that disables a model, tool, agent, integration, or class of side effects. |