Agentic Architecture & Orchestration

Concepts

• Agentic loop — resend-until-answered loop; only `tool_use` continues it.
• Stop reasons — `end_turn`/`tool_use`/`max_tokens`/`pause_turn`/`refusal`/`model_context_window_exceeded` and what each means.
• pause_turn handling — server-tool loop hit its cap → resend the conversation to continue.
• Tool execution types (client vs server) — you execute client tools; Anthropic runs server tools (bash/memory are still client-executed).
• Subagent decision criteria — spawn for parallel/isolated/independent work, not sequential/shared-state tasks.
• Adaptive thinking and effort — `type:'adaptive'` + `effort` (low→max); manual `budget_tokens` now errors on 4.6+.
• Interleaved thinking — reasoning between tool calls; automatic on 4.6+/Sonnet 4.6.
• Extended thinking with tool use — echo the full signed thinking block with its `tool_result`; `any`/`tool` incompatible.
• Task and Agent tool — spawns subagents; `Task`→`Agent` rename but `Task` persists in some surfaces.
• AgentDefinition — subagent identity/instructions/capabilities via `agents` param or `.claude/agents/*.md`.
• Hub-and-spoke orchestration — one orchestrator + specialist subagents in isolated context.
• Task decomposition (chaining vs adaptive) — static prompt chaining vs dynamic agent-decided decomposition.
• State across context windows — git checkpoints, framework-first, todo-driven iteration, memory tool.
• Workflow tool — TS Agent SDK tool to script orchestration of dozens–hundreds of agents.
• Workflows vs Agents — predefined code paths vs LLM-directed self-direction; start simple.
• Augmented LLM — the building block: LLM + retrieval + tools + memory.
• Prompt Chaining Workflow — fixed sequence of steps with optional "gates"; latency for accuracy.
• Routing Workflow — classify input → specialized handler; separation of concerns.
• Parallelization Workflow — run calls in parallel: sectioning (speed) vs voting (confidence).
• Orchestrator-Workers Workflow — orchestrator dynamically delegates *unpredictable* subtasks to workers.
• Evaluator-Optimizer Workflow — generator + evaluator loop against clear criteria.
• Autonomous Agents — LLM using tools in a loop; agents over workflows; ACI / poka-yoke tools.
• Architect's Reference Matrix
• Goal-oriented delegation
• PreToolUse hook enforcement
• Structured Intermediate Representations
• The Production Architecture Blueprint
• Zero-Tolerance Compliance