# Primary-source notes

Verified by opening the official pages on **17 October 2026**. These notes explain the conceptual basis and implemented local framework route. Actual pinned dependency execution is recorded in verification.md and the resolved environment receipt; no live model/service was exercised. The museum scenario, SVGs, examples, projects and exercises are original for this package. No source passage is reproduced at length.

| Source | Use in this package | Boundary |
| --- | --- | --- |
| [Anthropic, Building effective agents](https://www.anthropic.com/engineering/building-effective-agents), published 19 December 2024 | Distinguishes predetermined workflow control from model-directed agent control; informs the opening architecture question. | The page itself warns that tooling has changed. The course uses the conceptual distinction, not its model or product recommendations. |
| [OpenAI, Safety in building agents](https://developers.openai.com/api/docs/guides/agent-builder-safety) | Supports discussion of structured data boundaries, human tool confirmation and risks from untrusted text. | This is documented under the legacy Agent Builder area when checked. No Agent Builder product tutorial, model-version advice or security certification is provided here. |
| [AWS Builders' Library, Making retries safe with idempotent APIs](https://aws.amazon.com/builders-library/making-retries-safe-with-idempotent-APIs/) | Supports the distinction between caller-provided intent/identity, repeated delivery and safe handling of side effects. | The lab's serial in-memory replay ledger does not implement AWS-style durable atomic behavior. No AWS service is called. |
| [Python, unittest](https://docs.python.org/3/library/unittest.html) | Reference for standard-library tests, discovery and subtests. | Local execution used Python 3.14.4. The online `/3/` documentation resolved to 3.15 when checked; code does not depend on a 3.15-only feature. |
| [LangGraph overview](https://docs.langchain.com/oss/python/langgraph/overview) | Supports the primary-framework route and distinction between fixed and model-driven graph steps. | The executed migration uses fixed steps only; no model autonomy is measured. |
| [LangGraph interrupts](https://docs.langchain.com/oss/python/langgraph/interrupts) | Supports Module 7's actual interrupt/resume implementation and the warning that resumed node work executes again. | Tested locally with fixture decisions; no real approval service or persistent deployment. |
| [LangGraph persistence](https://docs.langchain.com/oss/python/langgraph/persistence) | Reference for graph checkpoints and thread identifiers. | The example uses InMemorySaver; it makes no cross-process durability claim. |
| [Python sqlite3](https://docs.python.org/3/library/sqlite3.html) | Supports the in-memory snapshot table, parameterized SQL and transaction context used in Modules 2–6. | Local single-connection tests do not establish distributed worker behavior or remote action atomicity. |
| [OpenAI evaluation best practices](https://developers.openai.com/api/docs/guides/evaluation-best-practices) | Reference for task-specific criteria and systematic evaluation. | Eight authored synthetic regressions are not a held-out model benchmark or population accuracy estimate. |
| [OWASP Prompt Injection](https://genai.owasp.org/llmrisk/llm01-prompt-injection/) | Describes risks when untrusted instructions enter model context; informs Module 9's future-adapter design exercise. | The executed project excludes raw notes from a deterministic planner, so it does not measure LLM attack resistance. |
| [LangGraph 1.2.16 on PyPI](https://pypi.org/project/langgraph/1.2.16/) | Primary package/version record for the pinned Module 7 dependency. | The exact installed dependency receipt is recorded locally; no environment is bundled or retained by this package. |

Before changing the framework dependency or adding any future live-model path, recheck official APIs and versions. Avoid copying vendor marketing claims as measured course outcomes. Record model, prompt, data, tool and dependency versions if a later permitted experiment actually executes them.