65 terms you should know
Agent SDK configuration object that specifies a subagent's description, system prompt, and tool restrictions. It defines each subagent type's role and boundaries within a multi-agent system.
The core control-flow cycle of an agent: send a request to Claude, inspect stop_reason, execute any requested tools, append the results to conversation history, and repeat. The loop continues while stop_reason is tool_use and terminates when it is end_turn.
Skill frontmatter field that restricts which tools may be used while a skill runs, for example limiting execution to file writes to prevent destructive actions.
Agent SDK configuration listing the tools an agent may call. It must include "Task" for a coordinator to spawn subagents, and is scoped narrowly to each agent's role to prevent cross-specialization misuse.
Degradation of review depth when too many files or too much content are analyzed in a single pass, producing inconsistent or even contradictory findings. It is mitigated by splitting work into focused per-file passes plus a separate integration pass.
Claude Code's native tools: Grep searches file contents for patterns, Glob matches file paths by name or extension, Read and Write handle whole files, Edit makes targeted changes via unique text matching, and Bash runs commands. Read plus Write is the fallback when Edit cannot find a unique anchor.
Continuous Integration and Continuous Deployment. Automated pipelines into which Claude Code is embedded using the -p flag and structured JSON output to run reviews, generate tests, and post feedback on pull requests.
Project directory holding custom slash commands that are shared with the whole team through version control. The user-scoped equivalent, ~/.claude/commands/, holds personal commands that are not shared.
The layered precedence of configuration files: user-level (~/.claude/CLAUDE.md), project-level (.claude/CLAUDE.md or a root CLAUDE.md), and directory-level files. User-level settings apply only to that user and are not shared with teammates through version control.
Directory of topic-specific rule files, each with YAML frontmatter. A paths field of glob patterns activates a rule only when a matching file is edited, so conventions load conditionally and consume tokens only when relevant.
Directory containing skills, each defined by a SKILL.md file. The frontmatter supports context: fork, allowed-tools, and argument-hint, and skills are invoked on demand for task-specific workflows rather than always loaded like CLAUDE.md.
Claude Code command that condenses the current conversation to reduce context usage during extended sessions that have filled with verbose discovery output.
Aligning a model's field-level confidence scores with actual accuracy using a labeled validation set, so review thresholds route the right extractions to humans. Self-reported confidence is otherwise poorly calibrated and unreliable as a complexity proxy.
Skill frontmatter option that runs the skill in an isolated sub-agent context, so verbose or exploratory output does not pollute the main conversation.
The central agent in a multi-agent system that decomposes the task, selects and invokes subagents, routes all inter-subagent communication, handles errors, and aggregates results. Overly narrow decomposition by the coordinator is a common cause of coverage gaps.
A topic area left unaddressed because task decomposition was too narrow or sources were unavailable. Synthesis output should annotate which findings are well-supported versus which areas have gaps rather than silently omitting them.
A caller-supplied identifier attached to each Message Batches API request. It correlates responses back to their requests and lets you resubmit only the specific documents that failed.
Making changes immediately without a planning phase. It is appropriate for simple, well-scoped tasks such as a single-file bug fix with a clear stack trace, in contrast to plan mode for architecturally significant work.
Use of ${VAR} references inside .mcp.json (for example ${GITHUB_TOKEN}) so authentication credentials are injected at runtime without committing secrets to the repository.
Specific, categorical rules stating exactly which issues to report versus skip (for example, flag a comment only when its claimed behavior contradicts the actual code). Explicit criteria improve precision far more than vague instructions like "be conservative" or "only report high-confidence findings."
A subagent used to isolate verbose discovery output during exploration and return only a summary, preserving the main conversation's context budget across multi-phase tasks.
Including a small set of targeted input/output examples (typically 2 to 4) in the prompt to achieve consistent formatting, demonstrate correct handling of ambiguous cases, and enable the model to generalize judgment to novel patterns rather than memorizing fixed cases.
Session mechanism that creates independent branches from a shared analysis baseline to explore divergent approaches, such as comparing two testing or refactoring strategies without redoing the initial codebase analysis.
A multi-agent architecture in which a central coordinator (the hub) mediates all communication with subagents (the spokes). Routing everything through the hub gives observability, consistent error handling, and controlled information flow.
CLAUDE.md syntax for referencing external files, keeping configuration modular. It lets each package's CLAUDE.md selectively include only the standards files relevant to its domain.
An iterative-refinement technique where Claude is asked to pose clarifying questions before implementing, surfacing considerations such as cache invalidation strategies or failure modes that the developer may not have anticipated.
The MCP flag a tool sets to signal a failure back to the agent. Effective error handling pairs isError with structured metadata (category, retryable flag, human-readable text) instead of a generic "Operation failed" message.
Claude Code CLI flag that constrains output to a supplied JSON schema. Used together with --output-format json so CI pipelines receive machine-parseable, schema-compliant findings that can be posted as inline PR comments.
A schema attached to a tool_use tool that guarantees syntactically valid, schema-compliant output. Strict schemas eliminate JSON syntax errors but do not prevent semantic errors such as line items that do not sum to the stated total.
The tendency of models to reliably use information at the beginning and end of a long input while omitting content in the middle. It is mitigated by placing key summaries at the start and organizing detailed results under explicit section headers.
The Claude API request parameter that caps how many tokens the model may generate in its response.
Project-scoped configuration file that declares shared MCP servers for a team via version control. It supports environment variable expansion so credentials are referenced without committing secrets.
An open standard for connecting Claude to external tools and data through MCP servers, exposing tools for actions and resources for content. It is one of the four core technologies in the exam scope.
A mechanism for exposing content catalogs, such as issue summaries, documentation hierarchies, or database schemas, so agents gain visibility into available data without making exploratory tool calls.
A process that exposes tools and resources over MCP. It is configured at project scope in .mcp.json for shared team tooling or at user scope in ~/.claude.json for personal servers, and tools from all configured servers are discovered at connection time and available simultaneously.
Claude Code command that shows which memory (CLAUDE.md) files are currently loaded, used to diagnose configuration hierarchy issues and inconsistent behavior across sessions.
An asynchronous Claude API offering roughly 50% cost savings with a processing window of up to 24 hours and no guaranteed latency SLA. It suits non-blocking, latency-tolerant jobs but does not support multi-turn tool calling within a single request.
Splitting a large review into per-file local-analysis passes plus a separate cross-file integration pass. This avoids attention dilution and the contradictory findings produced by analyzing many files at once.
Schema fields marked optional or nullable so the model returns null for information absent from a source document rather than fabricating a value to satisfy a required field.
Claude Code CLI flag that makes the tool emit results as JSON rather than free-form text, enabling automated pipelines to parse findings programmatically.
A Claude Code mode for safely exploring a codebase and designing an approach before making changes. It fits large-scale, multi-file, architecturally significant tasks with multiple valid approaches, preventing costly rework.
An Agent SDK hook that intercepts a tool's result after execution to transform or normalize it, for example unifying Unix timestamps, ISO 8601 strings, and numeric status codes before the model processes them.
Claude Code flag that runs the tool in non-interactive mode. It processes the prompt, writes the result to stdout, and exits without waiting for input, which prevents CI/CD jobs from hanging on interactive prompts.
Using hooks or prerequisite gates to guarantee workflow ordering deterministically, for example blocking process_refund until get_customer returns a verified customer ID. It is chosen over prompt instructions when business rules require guaranteed compliance, since prompts have a non-zero failure rate.
Iteratively condensing conversation history to save tokens. It is risky because it can drop precise numbers, percentages, dates, and customer-stated expectations, so transactional facts should be kept in a separate persistent block outside the summarized history.
Decomposing a task into a fixed sequence of focused passes, for example analyzing each file individually and then running a cross-file integration pass. It suits predictable multi-aspect workflows, whereas dynamic decomposition suits open-ended investigation.
Structured links between each claim and its source (URL, document name, excerpt, publication date) that subagents must preserve and merge through synthesis. Preserving these mappings prevents attribution from being lost during summarization steps.
A schema-validation library used to validate extracted output and drive validation-retry loops. It surfaces semantic validation errors (such as values in the wrong field) that strict tool-use schemas do not catch.
Claude Code CLI flag (--resume <session-name>) that continues a specific named prior conversation. Chosen when prior context is still mostly valid, whereas starting fresh with an injected summary is preferred when prior tool results are stale.
Giving each agent only the tools its role requires (for example 4 to 5 rather than 18) to keep tool selection reliable and prevent misuse of tools outside its specialization. Limited cross-role tools may be granted for specific high-frequency needs.
A file in which an agent records key findings so they persist across context boundaries. Referencing it for later questions counteracts context degradation during long exploration sessions.
Software Development Kit. In this exam it refers to the Claude Agent SDK, used to build agentic loops, spawn subagents via the Task tool, configure hooks, and manage sessions. It is one of the four core technologies in scope.
The observation that a model retains its generation reasoning within a session, making it less likely to question its own output. An independent review instance without that reasoning context catches more subtle issues than self-review instructions or extended thinking.
The principle that an independent Claude instance, lacking the generator's reasoning context, reviews code more effectively than the same session that produced it. It motivates using a separate instance for CI code review.
Service Level Agreement, a guarantee on latency or availability. The Message Batches API offers no latency SLA, which is why it is unsuitable for blocking workflows such as pre-merge checks.
The field on a Claude API response indicating why generation stopped. A value of tool_use means Claude is requesting tool execution and the agentic loop should continue, while end_turn means the turn is complete and the loop should terminate.
Stratified random sampling of high-confidence extractions across strata such as document type and field, used to measure ongoing error rates and detect novel error patterns that an aggregate accuracy figure can mask.
An MCP error payload that includes an errorCategory (transient, validation, permission), an isRetryable boolean, and human-readable text. It lets the agent make appropriate recovery decisions instead of retrying blindly or giving up on a generic failure.
A specialized agent spawned by a coordinator that operates with isolated context. It does not automatically inherit the parent's conversation history, so all needed context must be provided explicitly in its prompt.
The Agent SDK mechanism for spawning subagents. A coordinator must include "Task" in its allowedTools, and it can emit multiple Task calls in a single response to run subagents in parallel rather than across separate turns.
Using a hook to intercept an outgoing tool call and block a policy-violating action (for example a refund over $500), redirecting to an alternative workflow such as human escalation. It provides deterministic compliance that prompt instructions cannot.
The Claude API parameter controlling tool invocation. "auto" lets the model return text or call a tool, "any" forces the model to call some tool but lets it choose which, and forced selection ({"type": "tool", "name": "..."}) requires a specific named tool to be called.
The Claude API response content block in which Claude requests a tool call with structured input; it is also the stop_reason value that drives the agentic loop to execute a tool. Pairing tool_use with a JSON schema is the most reliable way to obtain schema-compliant structured output.
Error categories that shape agent recovery: transient errors (timeouts, unavailability) may be retried, validation errors indicate bad input, business errors indicate policy violations, and permission errors indicate access denials. Returning the correct category prevents wasted retry attempts.
A retry-with-error-feedback pattern that resends the original document, the failed extraction, and the specific validation errors so the model can self-correct. Retries succeed for format and structural errors but are ineffective when the required information is simply absent from the source.