# DeepThinking AI

> Independent, source-grounded analysis of AI systems: agent protocols, model benchmarks and the engineering behind them. Every claim cited.

Every article is available as raw markdown by appending `.md` to its URL.
Articles lead with a self-contained summary and list every source used.

## Agents & Protocols

How AI agents talk to tools, data and each other: MCP, tool calling, orchestration and the trust boundaries everyone forgets.

- [How do cacheable MCP tool lists change your token bill?](https://deepthinkingai.org/mcp-cacheable-tool-lists.md): MCP revision 2026-07-28 requires ttlMs and cacheScope on every list result and tells servers to return tools in a deterministic order. Both are filed as minor changes. They are the most direct cost change in the revision, because tool definitions sit in the cacheable prefix of every agent step. (published 2026-09-14)
- [Does MCP verify tool definitions the way it verifies skills?](https://deepthinkingai.org/mcp-skills-integrity-gap.md): SEP-2640 reached Final status and landed in the MCP docs on 13 September 2026. It requires hosts to verify every skill file against a SHA-256 digest and byte size, and binds approval to that manifest so one changed file revokes it. The tools primitive specifies none of this. MCP now has two content surfaces with opposite integrity models. (published 2026-09-14)
- [Progressive MCP vs CLI: which actually cuts tokens?](https://deepthinkingai.org/progressive-mcp-vs-cli.md): Anthropic measured one Drive to Salesforce transfer falling from 150,000 tokens to 2,000 by writing code against MCP servers instead of calling tools directly. That is a code execution result rather than an argument for deleting MCP. The Agent Skills spec puts numbers on the alternative: about 100 tokens per skill at rest. (published 2026-09-14)
- [How does the Model Context Protocol actually work?](https://deepthinkingai.org/how-model-context-protocol-works.md): The Model Context Protocol is a JSON-RPC 2.0 layer that lets an AI client discover and call capabilities exposed by separate servers. It defines three primitives: tools, resources and prompts. Since revision 2026-07-28 it is stateless, negotiating per request rather than per session. It standardises transport and discovery. Authorisation stays your job. (published 2026-08-19, updated 2026-09-14)

## Models & Benchmarks

What model releases actually change: context windows, reasoning modes, evaluation methodology and the gap between benchmark scores and real tasks.

- [What does a million-token context window actually buy you?](https://deepthinkingai.org/long-context-vs-retrieval.md): A large context window raises the ceiling on how much a model can see at once, but accuracy is not uniform across that window. Retrieval quality, latency and cost still dominate real workloads. Long context is best treated as a way to stop chunking small documents rather than as a replacement for retrieval. (published 2026-09-02)

## AI Engineering

The production craft: retrieval architecture, prompt caching, cost and latency budgets, evals and the operational realities of shipping LLM systems.

- [Do robots.txt and llms.txt stop AI agents from writing?](https://deepthinkingai.org/agent-controls-read-vs-write.md): robots.txt, llms.txt, Content-Signal and AIPREF all describe what an agent may fetch. The 2,000 packages uploaded to RubyGems in May went through a signup form and a publish endpoint, which none of those files govern. Read controls are advisory. Write paths need identity and rate limits. (published 2026-09-14)
- [When does prompt caching actually save money?](https://deepthinkingai.org/prompt-caching-economics.md): Prompt caching stores a prefix of your prompt so later requests reuse it instead of reprocessing it. Reads are far cheaper than base input tokens, but writing the cache costs a premium and entries expire. It pays whenever a large stable prefix is reused several times inside the TTL, and loses on one-shot traffic. (published 2026-09-11)

## Machine-readable endpoints

- [Search](https://deepthinkingai.org/search/?q={query}): substitute your terms for {query}. Runs the site's static index and returns matching articles. Also declared as a schema.org SearchAction.
- [What's new](https://deepthinkingai.org/whats-new.txt): dated changes, corrections and the current set of step-by-step guides. Re-fetch this to see what moved without re-reading the corpus.
- [Full text corpus](https://deepthinkingai.org/llms-full.txt): every article concatenated as markdown.
- [Posts index](https://deepthinkingai.org/api/posts.json): typed JSON with summaries, sources and markdown URLs.
- [JSON Feed](https://deepthinkingai.org/feed.json): JSON Feed 1.1.
- [RSS](https://deepthinkingai.org/rss.xml): RSS 2.0.

## About

- [Editorial methodology](https://deepthinkingai.org/methodology/): how articles are researched, sourced and updated.
- [About](https://deepthinkingai.org/about/)