Developer Foundations · 14.7% of the exam

agents and workflows

This is the first exam domain, about a seventh of the scored questions. It covers who decides an agent's next step, how an agent is built and where it runs, and the patterns and frameworks behind multi-step tasks.

Course example

Every domain follows Shelfwise, the course's own example rather than an exam scenario: staff at a small chain of hardware shops type a stock question into a chat widget, and an agent answers from the shops' stock feed. It reaches the feed through one small MCP server: a program that offers the feed as a tool over MCP, the Model Context Protocol. Three engineers run it as a long-lived Python service on the Agent SDK, the Anthropic library that runs Claude in a tool loop. Its targets are a mean cost under $0.01 an answer and 95% of answers within 8 s at the Saturday peak.

Domain 1 gives Shelfwise its agent design. It stays one agent after a coordinator with a branch-name subagent is priced and rejected (1.1). It is self-hosted, with a turn and budget limit and a lookup log written by an Agent SDK hook, a callback the service runs after successful stock lookups (1.2). Its only memory is each chat's own session, with auto memory turned off (1.3).

one agent, no subagentsself-hosted Agent SDK servicemax_turns=4 · max_budget_usd=0.03one session per chat
$0.01per-query cost ceiling, at about three requests an answer. A coordinator with a subagent would need 5 or more, and the subagent's requests would miss the cache. The $0.03 budget is about twice a cold answer, one whose first request writes the cache (≈$0.014).

Start hereThe words used throughout this domain

token
A short piece of text, often part of a word: the unit Claude reads, writes and is billed by, as input and output tokens.
context window
The most text, counted in tokens, that one request can hold, input and output together: the system prompt, the conversation so far, tool results and room for the reply.
prompt caching
Reusing the unchanged start of a request, its prefix, on later requests at a small fraction of the input price. Shelfwise caches a ≈6,000-token prefix; 2.3 shows how.
session
One agent conversation that the Agent SDK saves as it works, with an ID your code can pass back to resume it. Shelfwise starts one per staff chat (2.5).
setting_sources
The Agent SDK option that picks which Claude Code settings and instruction files a session loads. Left out, it loads the same ones as Claude Code. Shelfwise passes [], so none of them load, but auto memory still can (1.3).

task statements

videos

flashcards

0 / 9 known
Card 1 of 9