Tag: anthropic-sdk

A context window is everything the model can see in one request: system prompt, tool definitions, history, tool results, thinking and the reply. Here is what counts, what happens at the limit on claude-opus-5-5 and friends, and the six techniques agents use to work past it: deferred tools, context editing, compaction, memory, subagents and output filtering.

Tool calling is a protocol, not magic: the model emits a structured request matching a JSON Schema you supplied, your code runs it, and the result goes back as a message. The schema and its descriptions are rendered into the model's prompt and, with strict mode, into a decoding grammar, so they shape behaviour more reliably than any system-prompt instruction. Here is how it works on claude-sonnet-5-5 and OpenAI strict mode, with schema patterns that make invalid calls impossible.

In a workflow, your code decides the next step and the model fills in each step. In an agent, the model decides the next step and your code only executes it. That one question (who owns the control flow) determines cost, latency, testability and failure modes. Here is the distinction as Anthropic, Microsoft Agent Framework and LangGraph define it, with the same task built both ways in Python against claude-haiku-4-5 and claude-sonnet-5-5.

Every agent request pays for its system prompt, the provider's hidden tool-use preamble, and every tool schema before the user says anything. How to measure that fixed cost with differential count_tokens calls on the Anthropic API, the input_tokens endpoint on OpenAI, an MCP listTools probe, and /context in Claude Code, anchored to claude-opus-5-5, claude-sonnet-5, and @modelcontextprotocol/client 2.1.0.

Conversation history can live in your process, in your database, in the provider, or on the machine that ran the agent. Where you put it does not change your token bill, it decides your retention exposure, and it decides whether you can ever change providers. The numbers, the retention windows, and a storage shape that survives all three.

The 1M-token context window is now the default on Claude Opus 5, Opus 4.8/4.7/4.6, Sonnet 5, and Sonnet 4.6, with no beta header and no long-context premium. Here is when deleting the vector store actually pays off, the 10x cache-read rule that decides it, the ~30% tokenizer inflation that breaks your sizing estimate, and the seven-step migration with a verification line on each one.

A step-by-step checklist for retiring a hand-rolled while-loop agent and running the same job on @cursor/sdk 1.0.26. The tool bodies survive, the loop does not, and three things you probably rely on have no equivalent: a system prompt string, a tool allowlist, and per-turn control.

A step-by-step checklist for lifting a hand-rolled Anthropic tool-calling loop onto a standalone MCP server. The tool bodies move almost verbatim; what changes is where the schema comes from, how results are wrapped, and how errors cross the process boundary.

Use CodeAct (the agent writes executable code as its action) when your tasks chain many tools, loop, or move large data, and you can afford a sandbox. Use the JSON tool-calling loop for a handful of discrete, high-stakes actions where a code interpreter is overkill or unsafe. CodeAct wins on token cost and multi-step success rate; tool calling wins on safety and simplicity.

Use stdio for a local server one client launches, use Streamable HTTP for anything remote or multi-client, and do not build new HTTP+SSE servers -- that transport was deprecated in the 2025-03-26 MCP spec. Here is the decision, the wire-level differences, and the code for each.