We don't publish
your competitive advantage.

IFDeciding between a single API call, workflow, or agent architecture for a Claude-powered application.

THENStart with the simplest tier: single LLM call for one-shot tasks (classification, summarization, extraction, Q&A); workflow for multi-step code-orchestrated pipelines; Claude API + tool use for custom site_1 where you host compute; Managed site_2 for Anthropic-run agent loops with per-session containers, file mounts, and SSE event streams. Only escalate to a higher tier when task complexity, value, viability, and error-recoverability all justify it.

IFUser needs to decide between single API call, workflow, or agent architecture.

THENUse the decision tree: start with simplest tier (single call). For multi-step pipelines with code-controlled logic, use Claude API + tool use (workflow). For server-managed stateful site_1 with persistent configs and sandbox, use Managed site_2. Only choose agent if complexity, value, viability, and error tolerance are all satisfied.

IFDetermining which Claude surface (single API call, tool use, Managed site_1) to recommend for a new project

THENUse the decision tree: first check if deploying via third-party (Bedrock/Vertex/Foundry) → Claude API only. For a single LLM call (classification, extraction, Q&A) → Claude API. For multi-step pipelines with code-controlled logic → Claude API + tool use. For server-managed stateful site_1 with workspace, persisted configs, or long-running turns → Managed site_2. Default to the simplest tier that meets needs.

IFThe user is building an application with Claude and is unsure which interface tier to use (single call, workflow with tool use, or Managed site_1).

THENUse the decision tree: (1) If deploying on third-party (Bedrock/Vertex/Founded), Managed site_1 is unavailable — use Claude API + tool use. (2) For single LLM calls (classification, summarization, extraction, Q&A), use Claude API. (3) For multi-step pipelines with code-controlled logic, use Claude API + tool use. (4) For open-ended site_2 with your own tools and compute, use Claude API agentic loop. (5) Only use Managed site_3 when Anthropic should run the loop and host the tool-execution sandbox (e.g., stateful coding agent, long-running research agent, persisted agent configs). Check complexity, value, viability, and cost of error before choosing the agent tier.