Agentic Architecture & Orchestration

1. 1.1Agentic loopsBuild the loop that sends the conversation to Claude, runs the tools it requests, appends the results, and stops only when the stop reason is end_turn.
2. 1.2Coordinator-subagent orchestrationUse one coordinator to decompose a task, pick and delegate to specialist subagents, route all communication, and refine synthesis until coverage is sufficient.
3. 1.3Subagent invocation and context passingSpawn subagents with the Task tool, define each type with an AgentDefinition, and pass every needed finding explicitly as structured data because subagents start with no context.
4. 1.4Workflow enforcement and handoffEnforce must-always-happen steps with hooks or prerequisite gates in code, decompose multi-concern requests, and hand off to humans with a structured summary.
5. 1.5Agent SDK hooksUse PostToolUse hooks to normalize tool results and interception hooks to block policy-violating calls, choosing hooks over prompts when compliance must be guaranteed.
6. 1.6Task decomposition strategiesChoose prompt chaining for predictable multi-aspect work and adaptive decomposition for open-ended investigation, splitting large reviews into focused passes.
7. 1.7Session state, resumption and forkingResume named sessions, fork a shared analysis to compare approaches, flag changed files on resume, and start fresh with a summary when prior results are stale.

• Keep the loop going while the stop reason is tool_use and end it on end_turn, never on what the reply text says.
• Let a coordinator pick only the subagents a request needs instead of running the full pipeline every time.
• Pass a subagent everything it needs in its prompt; it starts with no memory of the coordinator's conversation.
• Fire independent subagents in parallel from a single coordinator response.
• When a step must always happen (identity check before a refund), enforce it in code with a hook or gate, not with prompt wording.
• Use prompt chaining when the steps are known in advance, and adaptive decomposition when the work is an open investigation.
• Resume a session when earlier context is still accurate; start fresh with a summary when the code underneath has changed a lot.

• Ending the loop when the reply contains text, or relying on a fixed iteration cap as the main stop.
• Assuming subagents automatically share the coordinator's history or each other's results.
• Subagents that talk to each other directly, bypassing the coordinator.
• Writing a stronger system prompt to guarantee a compliance rule.
• Escalating to a human with a vague note instead of a structured handoff (customer, root cause, recommended action).

You get 4 of the 6 scenarios at random, so practice this domain in each of them. All six scenarios.

1. Write a minimal agent loop with two tools and log the stop reason on every pass.
2. Build a coordinator with two subagents and pass findings between them as structured data with sources attached.
3. Add a hook that blocks a tool call until a prerequisite tool has returned a verified result.
4. Resume a named session after changing a file, and practice telling the agent exactly what changed.