# Resources & Sources > All sources used for this lesson, organized by tier. > Every claim in lesson.agent.md traces back to one of these. --- ## Tier 1: Official Documentation > Primary sources. Check these first. - **Anthropic — Tool Use (Function Calling) Overview** — `sources/official-docs/anthropic-tool-use.md` - URL: https://platform.claude.com/docs/en/docs/build-with-claude/tool-use/overview - What to read: tools are `name` + `description` + `input_schema`; the model returns `stop_reason: "tool_use"` with `tool_use` blocks; you execute and return a `tool_result`; the **agentic loop** repeats until it stops; `tool_choice` (auto/any/tool/none); tools as structured output. This is the protocol the project implements. - **Anthropic — Building Effective Agents** — `sources/articles/building-effective-agents.md` - URL: https://www.anthropic.com/engineering/building-effective-agents - What to read: the **definition of an agent** ("LLMs using tools based on environmental feedback in a loop"), workflows vs. agents, the **augmented LLM** (retrieval + tools + memory), and the **agent-computer interface (ACI)** — invest in tool descriptions. First-party engineering guidance. - **LiteLLM — completion()** — `sources/official-docs/litellm-completion.md` (carried from Project 01) - URL: https://docs.litellm.ai/docs/completion/input - What to read: the `tools` / `tool_choice` params and the OpenAI-style `tool_calls` response shape — each call's `function.arguments` is a **JSON string** you must `json.loads`. The lab's cross-provider tool path. --- ## Tier 2: Foundational Papers > Academic papers that established the concepts in this lesson. - **ReAct** — `sources/papers/react-paper.md` - Title: ReAct: Synergizing Reasoning and Acting in Language Models - Authors: Yao, Zhao, Yu, Du, Shafran, Narasimhan, Cao - Year: 2022 - URL: https://arxiv.org/abs/2210.03629 - Why it matters: the foundational pattern of the loop — **interleaving reasoning traces and actions** (Thought → Action → Observation). Acting fetches ground truth, which is why it "overcomes issues of hallucination" present in reasoning-only prompting. The conceptual ancestor of tool-use APIs. - **RAG** — `sources/papers/rag-paper.md` (carried from Project 04) - URL: https://arxiv.org/abs/2005.11401 - Why it matters: retrieval-augmented generation — the **context-injection** half (retrieve relevant files, ground the answer in them) the copilot reuses before reaching for tools. --- ## Tier 3: Engineering Guides > Practical engineering perspectives. - **Chunking Strategies** — `sources/articles/chunking-strategies.md` (carried from Project 03/04) - URL: https://www.pinecone.io/learn/chunking-strategies/ - Why it matters: for the extended "chunk + embed per-chunk" code retrieval — chunk-size precision/context tradeoff and overlap, applied to source files. --- ## Tier 4: Educational Sources > Tutorials, courses, and explanatory content useful for learning. - (None specific to this lesson. The Anthropic tool-use docs and the ReAct paper are themselves the primary teaching sources; the Project 03/04 retrieval sources are the conceptual prerequisites for context injection.) --- ## Optional — Going Deeper (MCP) > Read **after** the lab works and you can explain your own tool loop. The Model Context Protocol is > what that hand-wired loop becomes when many hosts must reuse the same tools over a wire protocol. > These reframe the project; they are not required to complete it. The **MCP Interface Layer** > elective builds on them directly. - **MCP — Architecture** — `sources/official-docs/mcp-architecture.md` *(optional — depth)* - URL: https://modelcontextprotocol.io/specification/2025-06-18/architecture - Why it matters: the client-host-server split — a host coordinates clients, each client talks to one focused server. Your monolithic copilot loop, decomposed into a reusable protocol surface. - Where you'll see this: any editor/IDE or chat host (Claude Desktop, Cursor) that loads tools from external servers instead of hard-coding them. - **MCP — Tools** — `sources/official-docs/mcp-tools.md` *(optional — depth)* - URL: https://modelcontextprotocol.io/specification/2025-06-18/server/tools - Why it matters: tool schema and `tool_result` as a standardized wire format — the same `name` + `description` + `input_schema` you wrote, but provider- and host-agnostic. - Where you'll see this: every MCP server that exposes a tool catalog a host can discover and call without bespoke glue code. - **MCP — Security Best Practices** — `sources/official-docs/mcp-security-best-practices.md` *(optional — depth)* - URL: https://modelcontextprotocol.io/specification/2025-06-18/basic/security_best_practices - Why it matters: trust boundaries, least privilege, and local-server risk — your `safe_resolve` containment check generalized into protocol-level guidance. - Where you'll see this: hardening any tool server that touches the filesystem, shell, or network on a user's machine. --- ## Recommended Reading Order For a learner new to this topic: 1. Start with: `sources/articles/building-effective-agents.md` — what an agent *is* (an LLM using tools in a loop) and why to keep it simple. 2. Then read: `sources/official-docs/anthropic-tool-use.md` — the concrete protocol (schema → call → result → loop) and `tool_choice`. 3. Then read: `sources/papers/react-paper.md` — why interleaving reasoning and acting reduces hallucination and makes the loop terminate. 4. Reference: `sources/official-docs/litellm-completion.md` — the exact `tools` / `tool_calls` shape the code uses (and the JSON-string `arguments` trap). 5. Reference: `sources/papers/rag-paper.md` — the context-injection half (carried from Project 04). --- ## Further Reading Topics adjacent to this lesson worth exploring later: - Self-editing memory as tool use: letting the model decide what to remember/retrieve (MemGPT; bridge from Project 05 to Project 08). - Multi-tool planning, orchestrator-workers, and evaluator-optimizer agent patterns (Building Effective Agents; bridge to Project 08). - Strict/structured tool output and schema validation for reliable final answers (Anthropic tool-use docs). - Sandboxing and permissioning real tools (filesystem, shell, network) — the security surface of an acting agent. - Evaluating tool-using agents: correctness, tool-call accuracy, and hallucination over actions (bridge to Project 07).