Study notes · 4.9% of the exam

Agent Patterns and Frameworks

Apply the core agent patterns (tool-use loops, subagents, memory, context-window management) to multi-step tasks, and reason about when and how agent frameworks such as Strands, LangGraph and PydanticAI help.

Key points

1
Tool-use loop: Claude requests tools (stop_reason: "tool_use"), your code runs them and returns results, and the loop repeats. A pause_turn stop means a server-tool loop paused; send the assistant content back to continue.
2
Claude may call several tools in one response. Return all results together, matched by tool_use_id. If Claude must see each result before choosing the next step, set disable_parallel_tool_use: true inside tool_choice.
3
Return tool failures as tool_result with is_error: true and an actionable message. A bare "error" gives the model nothing to change and often causes repeated identical calls.
4
Subagents start with a fresh conversation, not the parent's history, and only their final response returns to the parent. Pass file paths, errors and decisions in the delegation prompt or in a notes file.
5
The model keeps no state between requests, so memory is an application pattern. The memory tool is client-side: Claude requests file operations under /memories, and your app executes them in storage you control (rejecting paths outside that directory).
6
Context is a finite attention budget, and quality degrades as it fills with irrelevant tokens ("context rot"). Aim for the smallest set of high-signal tokens; a bigger context window does not cure context pollution.
7
Context editing, specifically tool-result clearing, drops old tool results (leaving a placeholder) and keeps the rest verbatim. Compaction summarizes older history and is the broader default strategy for long conversations.
8
Just-in-time retrieval: keep lightweight references (paths, queries) and load data with tools when needed. This avoids stale indexes on fast-changing sources. A hybrid approach retrieves some data up front.
9
Compaction can drop instructions given early in a conversation. Keep persistent rules in CLAUDE.md, enforce must-hold rules with hooks, and use a PreCompact hook to archive transcripts if needed.
10
Long-horizon work across context windows: an initializer session writes a feature list, a progress file and a first commit. Later sessions take one feature at a time, test it end to end, update progress and commit, so each fresh window can rebuild the state.
11
LangGraph is a low-level orchestration framework for long-running, stateful agents. It builds graphs of nodes and edges over shared state (StateGraph) and supports durable execution, human-in-the-loop interrupts and memory.
12
PydanticAI is a typed Python agent framework: output types and tools are validated with Pydantic models, and dependencies are injected through a typed run context. Strands Agents is an open-source, model-driven agent SDK (Python and TypeScript) with hooks, multi-agent patterns and multiple model providers. All three can use Claude.
13
Framework guidance: start with the API directly, since many patterns take only a few lines. Frameworks add abstraction that can obscure prompts and responses, so if quality changes after adopting one, inspect the actual requests it sends.

Read the source

Test yourself on Agent Patterns and Frameworks

Ten questions, with the answer and explanation after each one.