THE CORE AIM OF VOCABULARY MASTERY · AI AGENT DEPENDENCY MAPPING · DISCOVER → CONNECT → TEST → MONITOR → REPAIR
When a student opens an online worksheet, everything looks simple: a question, a place to type and an answer. But the response may depend on login, a learner record, an AI model, a search index, a curriculum reference, a tool server and a network connection. If one part fails, another part may behave differently. AI agent dependency mapping vocabulary helps readers see those hidden relationships.
The core words are upstream dependency, downstream consumer, dependency graph, direct dependency, transitive dependency, critical path, single point of failure, shared dependency, failure propagation, blast radius, fallback and dependency owner. The core aim is system-relationship clarity: identify what relies on what, decide which connections are essential, predict which users may be affected by a change, and distinguish failure of a supporting service from failure of the educational idea itself.
This guide is part of the eduKateSG Vocabulary Hub. It differs from AI Agent Supply-Chain Integrity Vocabulary: an SBOM helps identify installed software components, while dependency mapping describes the wider operational and informational relationships among tools, sources, people, records and workflows. The connection to AI Agent Orchestration Vocabulary is equally important: orchestration decides how work is coordinated; a dependency map explains what that work depends upon.
A parent-facing example: one broken source, many confusing answers
Imagine the current school syllabus is available from an official public document, and an educational assistant indexes it for search. A student asks about a Secondary 1 topic. The tool can still answer from its cached index after the official PDF changes. Technically, the model and website are working. But the source freshness dependency has failed. An operator who maps only computers and network connections could miss that educational dependency.
A useful map therefore includes both technical services and factual inputs. For example: official document → extraction → indexed passage → retrieved evidence → answer. If the upstream document is revised, downstream cached content needs review. Not every old answer becomes false, but all relevant claims must be checked for applicability.
The dependency mapping glossary
| Term | Meaning | What to ask |
|---|---|---|
| Dependency | A component or condition required for another operation. | What does this step rely on? |
| Upstream | A supplier of data or functionality to another component. | What feeds this step? |
| Downstream | A consumer affected by an earlier component’s output. | Who relies on this result? |
| Direct dependency | A directly used component or service. | What does the application call itself? |
| Transitive dependency | A dependency reached through another dependency. | What sits further along the chain? |
| Dependency graph | A representation of nodes and their relationships. | How are the pieces linked? |
| Node | One component, source, process or role in a graph. | What is being represented? |
| Edge | A recorded relationship between nodes. | What relies on what? |
| Critical path | The chain of essential steps that determines completion timing in a workflow. | Which steps determine the finish? |
| Shared dependency | One underlying service used by multiple systems. | What could fail together? |
| Single point of failure | A component whose loss can disable a larger function. | What has no usable substitute? |
| Dependency owner | The person or team responsible for a particular component. | Who investigates or changes it? |
| Dependency contract | Agreed inputs, outputs and availability expectations. | What does this component promise? |
| Availability dependency | A requirement that another service be reachable. | What if it goes offline? |
| Data dependency | Reliance on a dataset, source or record. | What if the information changes? |
| Permission dependency | Reliance on an access grant or scope. | What if a credential is revoked? |
| Failure propagation | A problem spreading from one component into its consumers. | What breaks next? |
| Blast radius | The scope of systems or users affected by a problem. | How large is the impact? |
| Fallback | A permitted alternative when a dependency cannot be used. | How can the task continue safely? |
| Observability | Signals that reveal how dependencies are behaving. | What evidence shows a failure? |
| Circuit breaker | A mechanism that temporarily limits calls to unhealthy services. | Should requests be paused? |
| Dependency inventory | An organised list of key relationships and owners. | What have we mapped? |
The direction of arrows matters
Suppose the article generator relies on a vocabulary source registry. We can draw ‘article generator → source registry’ to mean the generator depends on the registry. But another team might draw the arrow from the source toward its consumers to show information flow. Either convention can work if it is stated clearly. Without a legend, one reader may interpret an arrow backwards and investigate the wrong system.
A second ambiguity involves the word blocked. A source may be reachable but legally unsuitable for redistribution. A tool may be running but no longer authorised for this student. Maps should label what kind of dependency is represented—availability, data validity, permission, licence, version or execution order—rather than putting every connection in one undifferentiated line.
A complete educational AI graph has several layers
| Layer | Example node | Why it belongs |
|---|---|---|
| User and identity | Authenticated parent or learner | Controls relevant access and context. |
| Curriculum | Current official syllabus edition | Anchors formal learning requirements. |
| Content | Reviewed concept glossary | Supplies definitions and examples. |
| Model | Language model configured for explanation | Generates readable responses. |
| Tools | Search, calculator or WordPress connector | Performs external operations. |
| State | Current learner step or saved task progress | Prevents incorrect next actions. |
| Infrastructure | Database, index and network service | Provides runtime capacity. |
| Review | Teacher or source-verification gate | Checks consequential outputs. |
| Destination | Published post, worksheet or message | Receives the actual final result. |
Worked example: four dependent steps in a vocabulary article
An editor asks an assistant to publish a new article on the word inference. Before publication, the writer needs current reference material, the editor needs a draft, the source reviewer needs the specific claims and the publisher needs the reviewed version plus authority to write. The chain is not a single action. Every dependency has its own failure and a different repair.
| Stage | Required dependency | What failure would look like |
|---|---|---|
| Research | Source library and applicable passages | Plausible but unsupported definition. |
| Draft | Research findings and intended audience | Generic or wrong-level explanation. |
| Review | Exact draft version and factual sources | Old review applied to new text. |
| Publish | Authorised WordPress tool and reviewed body | Draft mistaken for live article. |
| Verify | Actual post ID, status and saved content | Invented or broken public link. |
If WordPress returns a failure, rerunning source research is unlikely to fix the publishing dependency. If the official source is outdated, retrying the WordPress call is the wrong repair. The map allows the operator to address the failure at its actual location instead of trying arbitrary model prompts.
Shared dependencies create hidden blast radius
Two different eduKate services might rely on one central reference index. One index update could affect English vocabulary lessons, science explainers and mathematics terminology at the same time. Looking at any one site alone may not reveal why several branches suddenly have similar citation problems.
A dependency graph can identify every known downstream consumer of the shared service. If a bug appears in that service, the operations team can prioritise the affected pages and assessments. Because the inventory may be incomplete, monitoring and user feedback still matter; unmapped consumers can exist.
A single point of failure is not always a machine
Sometimes the single point is a sole reviewer, an expiring credential, a manual source-update procedure or an external provider with no accepted alternative. For example, if only one person knows which school-year syllabus source is authoritative, the system may become unreliable when that person is unavailable.
The repair might be documentation, a second authorised reviewer, a source registry or a defined contingency process. Buying another server will not repair a missing governance dependency. This is why agent dependency mapping belongs to both technical and organisational management.
How SRE monitoring supports dependency awareness
Google’s Site Reliability Engineering workbook recommends monitoring responses from direct dependencies, including their latency, error codes and relevant resource signals. This makes sense because a service can appear healthy while its dependent service becomes slow or unreliable. For AI workflows, a similar approach can monitor tool results, source lookups, document indexing and downstream write operations.
The monitoring must also distinguish the technical response from factual usefulness. A successful source lookup may return a page that no longer applies to the current school year. Instrumentation can show the lookup succeeded; source validation must assess the content.
Dependency map versus task graph versus source graph
| Graph type | Main question | Example |
|---|---|---|
| Dependency graph | What relies on what? | Article generation relies on the glossary index. |
| Task graph | In what order should work happen? | Research before review, review before publication. |
| Source-provenance graph | Where did a claim originate? | Explanation was derived from a specific document passage. |
| Organisational ownership map | Who maintains or approves a component? | Science source reviewer is responsible for syllabus applicability. |
| Permission graph | Which actors may access resources? | A parent may read their own child’s progress. |
These graphs can be connected but are not interchangeable. A task may depend on a particular data source even when the source is not itself a step the agent performs. A person may own a system without having direct access to every private record it contains.
A dependency mapping checklist
- State what a graph arrow means.
- Inventory important tools, datasets, indexes, models, accounts and external services.
- Record component owners and known contract versions.
- Label relationships by data, availability, permission and execution requirements.
- Identify shared dependencies and potential correlated failures.
- Find critical dependencies with no permitted fallback.
- Record which educational outputs might use stale sources.
- Check how external publication and delivery state will be verified.
- Maintain change history when a dependency is upgraded or retired.
- Treat unknown or unmapped portions explicitly as gaps rather than assuming they are absent.
An exercise for students: trace a failing tutor
A student can still open the webpage, but the AI cannot retrieve practice questions. List three candidates: the model, the practice-question index and the student’s authentication scope. What evidence would distinguish them? A search tool error suggests the index or permission path; a model-generation error points elsewhere; a valid search with an irrelevant year-level result suggests a routing or source-validity problem.
Now draw the path from the official curriculum reference to the final practice question. Mark every step that would require another check if the official document changed. The exercise teaches that continuity is preserved by connections, not merely by storing more documents.
Common dependency mapping mistakes
- Only listing installed packages: include services, data and operational ownership where relevant.
- Unlabelled arrows: document relationship direction and type.
- Assuming availability equals correctness: check source applicability.
- Ignoring shared dependencies: examine correlated failures across products.
- Thinking every dependency is equally critical: assess impact and fallback.
- Automatically trusting a backup: verify compatibility and permissions.
- Mapping only the happy path: include approval, rejection and recovery routes.
- Assuming the diagram is complete: record confidence and missing coverage.
Frequently asked questions
What is AI agent dependency mapping?
It is the representation of services, information, permissions and other elements on which agent workflows rely, including their relationships and potential failure effects.
Is a dependency map the same as an SBOM?
No. An SBOM focuses on software component inventories; a broader agent dependency map can include data sources, model services, people and permissions.
What is a downstream consumer?
A component or user-facing system whose work depends on the output or availability of another component.
What is a shared dependency?
One underlying service or source relied on by multiple systems, creating a possible common failure point.
Why map educational sources?
Because an outdated or incorrectly linked source can affect downstream explanations even when technical services remain operational.
References and connected learning
- Google SRE Workbook — Monitoring (dependency monitoring and operational signals).
- OWASP Dependency Graph and SBOM guidance
- NIST AI RMF Core (mapping components and risks).
- AI Agent Supply-Chain Integrity Vocabulary
- AI Agent Observability Vocabulary
- AI Workflow Recovery Vocabulary
- AI Source Verification Vocabulary
The AI agent dependency mapping vocabulary standard
Mastery means being able to draw and explain a dependency chain, identify upstream and downstream roles, distinguish operational from informational failure, locate shared dependencies and propose an appropriate verification step when something changes. This is how an AI teaching system avoids treating every confusing answer as simply a failure of the language model.