Foundations domain 1 · Agentic architecture and orchestration

27% of the Foundations exam · Last checked: 30 September 2026

The agentic loop and stop reasons

A loop sends a request, reads why the model stopped, and acts on that. The stop_reason is the signal. Don't parse the text to guess.

stop_reason	Meaning	What to do
`end_turn`	The model finished	Exit the loop
`tool_use`	It wants a tool run	Run it, send the result back, loop again
`max_tokens`	Output hit your limit and was cut off	Treat as unfinished. If it was cut off inside a tool call, the input is incomplete: don't execute it; retry with more room
`stop_sequence`	One of your stop strings was produced	Check which one, then usually exit
`pause_turn`	A long-running turn (such as server tools) was paused	Send the response back as it is to let it continue
`refusal`	The model declined	Don't resend the same prompt unchanged
`model_context_window_exceeded`	The context window filled up	Reduce or restructure what you send

Exam guide vs todayStudy notes often list six stop reasons. Current docs have a seventh, model_context_window_exceeded. On the newest models, a refused request can sometimes be served by retrying on a different Claude model. The exam guide itself names only tool_use and end_turn.

Always pair the loop with a safety bound, such as a maximum number of iterations or a budget, and make hitting it a visible outcome, not a silent stop.

Tool results and parallel calls

Tool results go back in a user-role message. "User" is the side of the conversation that isn't the model; no human has to be involved.
When the model requests several tools in one response, return all results in one user message, one tool_result block per call, matched by tool_use_id. The tool_result blocks come before any text.
A failed call still gets a result: a tool_result with is_error: true and a useful message. An unanswered call is a malformed conversation.

Hooks: deterministic enforcement

PreToolUse runs before a tool and can block it. PostToolUse runs after, so it can't prevent anything. It can inspect, normalise or replace the output, or add context.
Exit codes: 0 success; 2 blocking error (for PreToolUse, the call is blocked and stderr goes back to Claude); any other non-zero code is a non-blocking error and the call proceeds. If a hook prints valid JSON, the JSON decision is used.
The principle behind the facts: prompts are probabilistic, hooks and permissions are not. "Must never", "zero tolerance", "regulatory", or "a prompt rule already failed" point to a hook or permission.

Exam guide vs todayOlder notes list nine hook events: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Notification, Stop, SubagentStop, PreCompact, SessionEnd. Those are still the core, but Claude Code now has more than thirty. Examples include PermissionRequest, PostToolUseFailure, SubagentStart, PostCompact and InstructionsLoaded.

Subagents and coordination

A subagent starts without the coordinator's conversation. Everything it needs has to be in its prompt: the goal, the inputs, what a good answer looks like and the output format. (A fork is the exception: it inherits the whole conversation.)
Give each subagent only the tools its role needs. A subagent defined without a tools list inherits all of them.
Claude decides when to delegate to a subagent from its description, so a vague description means it isn't used.
Hub-and-spoke: subagents report to the coordinator, which owns delegation, keeps it bounded, and synthesises. Independent subtasks can run in parallel.
Open-ended investigation needs adaptive decomposition: plan, then revise as findings come in. Fixed, repeated processes suit a fixed workflow.

Sessions

claude --continue resumes the most recent conversation in this directory. claude --resume lets you pick one.
Forking a session copies its history into a new session to try an alternative. Forks don't merge back.
If much of a session's earlier tool output is stale, start a fresh session with a structured summary instead of resuming.

Docs: Stop reasons Handling tool calls Parallel tool use Hooks Subagents Sessions

Practise Domain 1