NOTE

Literature: Anthropic Advanced Capabilities

authorclaude-sonnet-4-6 aliases source00_Raw/anthropic/ titleLiterature: Anthropic Advanced Capabilities statusactive date2026-05-02 typeliterature

Literature: Anthropic Advanced Capabilities

Synthesis of the second-batch Anthropic raw corpus: advanced reasoning modes, batch execution, file management, tool extensions, and the Managed Agents surface. The first-batch fundamentals (Messages API, tool use, streaming, error handling, prompt caching basics) are covered in lit-anthropic-messages-api.

Source Set

Crawled from platform.claude.com/docs between 2026-05-01 and 2026-05-03:

File Topic
adaptive-thinking.md Adaptive thinking mode, effort parameter, display control
extended-thinking.md Manual thinking mode, interleaved thinking, block preservation
effort-parameter.md effort levels, model-specific guidance, tool use interaction
batch-processing.md Message Batches API, polling, extended output beta
files-api.md Files API, upload/reference pattern, storage lifecycle
token-counting-api.md Token count endpoint, pre-call estimation
tool-use-mcp-connector.md MCP Connector, toolset config, allowlist/denylist, TypeScript helpers
tool-use-runner-sdk.md Tool Runner SDK, @beta_tool, compaction
tool-use-server-tools.md Server-side tool types
managed-agents-quickstart.md Agent/Environment/Session model, quickstart
managed-agents-sessions.md Session lifecycle, statuses, vault IDs
managed-agents-agent-setup.md Agent configuration and versioning
managed-agents-environments.md Environment templates
managed-agents-events-streaming.md Event types, streaming protocol
managed-agents-tools.md Managed agent toolset
models-overview.md Model IDs, token limits, feature support matrix
models-api-reference.md Model API reference
api-beta-headers.md Beta feature headers
api-versioning.md API version policy
context-editing.md Context compaction, thinking block clearing
client-sdks.md SDK overview

Durable Architectural Patterns

Adaptive thinking replaces manual token budgets. The shift from budget_tokens to effort is not cosmetic — it changes who controls thinking depth (model vs. caller). Adaptive mode outperforms fixed budgets on bimodal tasks and long-horizon agentic workflows because Claude can skip thinking for simple sub-tasks. Manual mode gives predictability; adaptive mode gives optimality.

Interleaved thinking is automatic in adaptive mode. Claude reasons between tool calls, not just before the first one. This is architecturally significant for multi-step tool workflows: reasoning quality at each step improves when Claude can reconsider after seeing a tool result.

Thinking display is a latency knob, not a cost knob. display: "omitted" skips streaming thinking tokens, reducing time-to-first-text-token. You are charged the same for full thinking tokens regardless of display setting.

Batch API = cost-optimized async tier. 50% discount for any workload that can tolerate < 1 hour latency. Batch and prompt caching discounts stack. The extended output beta (300k tokens) is batch-only — synchronous API caps at 64k–128k depending on model.

Files API decouples ingestion from inference. The upload-once pattern eliminates repeated base64 encoding of large documents. The asymmetry — uploads cannot be downloaded, only code-execution outputs can — reflects a specific design intent: files are inputs to inference, not a general object store.

MCP Connector makes Claude the MCP client. The simplification is real: no MCP client implementation, no local tool execution infrastructure for remote MCP tools. The constraint is also real: only HTTP-accessible servers, only tool calls (not prompts or resources), not on Bedrock or Vertex AI.

Tool Runner SDK is loop automation, not a protocol change. The same tool call loop still runs; the SDK handles the iteration, error wrapping, and state management. Compaction support is the most important addition for production agentic workloads — it allows agents to run beyond context window limits.

Managed Agents is Anthropic's hosted agent runtime. Compared to building on the Messages API, the tradeoff is control vs. simplicity: container provisioning, agent loop, and tool execution all happen server-side. Sessions are event-driven state machines; the caller sends user events and receives agent events via SSE.

Operational Details Likely to Drift

  • Beta headers (files-api, mcp-client, managed-agents) will change as features GA.
  • mcp-client-2025-04-04 is already deprecated; mcp-client-2025-11-20 is current.
  • output-300k-2026-03-24 beta header for extended batch output — likely to be versioned.
  • Model-specific effort level availability (xhigh is Opus 4.7-only as of this capture).
  • Exact pricing tables (batch rates, cache write/read rates) — treat the pricing page as authoritative.
  • Files API rate limit (100 req/min) is a beta-period limit and will change.
  • Managed Agents tool types (agent_toolset_20260401) use date-versioned identifiers.

Relationship to First Batch

The first batch (lit-anthropic-messages-api) covers: Messages API request/response shape, tool use execution loop, SSE streaming baseline, error handling, prompt caching core mechanics.

This batch extends it with: reasoning modes, async execution, file storage, MCP integration, SDK-level abstractions, and the hosted agent runtime. The two batches together cover the full current Anthropic API surface captured in the local corpus.

See also