Sources, every one with its date:

• Our own three arm test of the ignore file that does not work — AI University research note, run 2026-08-21 on Claude Code 2.1.236. A junk file containing a canary string, read back by a headless agent under three conditions. Arm one, no ignore file: read succeeded. Arm two, an ignore file naming the junk file exactly: read succeeded, canary returned verbatim. Arm three, a deny rule on reading that path in settings, with the ignore file still present: read denied. The method is here so you can rerun it rather than take our word for it. The working mechanism is a deny rule under permissions in your settings file, documented at the settings link in these notes
• The Register on Claude asserting a protection feature that does not exist — Thomas Claburn, 2026-01-28. Reports Claude telling a user that an ignore file works like gitignore and that Claude Code will refuse to read matching files. The article's verdict is flat: Claude is incorrect. Anthropic did not respond to a request for comment. This matters because people believed their environment files were protected
• JetBrains measured a terse prompting tool on real agentic work — Denis Shiryaev, July 2026. Claude Code headless, claude-sonnet-5 at low effort, SkillsBench, 86 tasks and 82 clean pairs, roughly 240 trials. The tool claims about sixty five percent output reduction on its own prose benchmark. Measured on coding: output tokens down 8.5 percent, cost about ten percent, quality indistinguishable at p equals 0.82. Their verdict: safe, honest about style, oversold on savings. Note the tool's readme now carries both numbers and explains the gap
• JetBrains measured the sixty to ninety percent claim, and it came out negative — Same author and rig. Four paired runs, 425 billed trials, about 320 dollars. The tool's readme promises sixty to ninety percent less token consumption. Measured: a median 7.6 percent increase in cost per task at low effort, with quality a wash. Their explanation for why the claim had nowhere to live is that the hook only ever sees about a fifth of the tool output, and most input is cached re reads billed at the lower rate
• The bug report the eighteen thousand token figure actually comes from — Opened 2025-07-12. Read the numbers carefully: built in tools alone are ten to eleven thousand tokens, MCP servers add four to eight thousand on top, and the total overhead is ten to twenty thousand tokens at conversation start. It is a TOTAL for built ins plus MCP, not a per server figure, and the reproduction rows make that explicit
• Per server MCP costs, actually measured, and they vary twenty five fold — A benchmark that states its method: stdio handshake, list the tools, serialise name, description and input schema, count with a named tokenizer. Slack comes in around 679 tokens. Notion comes in around 17,161 for 24 tools. So one server costing eighteen thousand tokens is true of the worst one and about twenty five times wrong for a typical one, and nobody repeating the number says which
• The cache lifetime that moved twice without an announcement — A user's analysis of 119,866 API calls across two machines, January to April 2026, reading raw session logs. Documents the one hour cache lifetime silently reverting to five minutes around 6 March 2026, a twenty to thirty two percent rise in cache creation cost, and subscription users hitting quota limits for the first time. Closed as not planned with no vendor reply. This is why advice that looks wrong today was often right when it was written
• How Claude Code uses prompt caching, including every way you break it — The primary source for the eight breakers and the eight safe actions. Switching models, changing effort, fast mode, connecting or disconnecting a server whose tools are loaded upfront, toggling a plugin, denying a tool by bare name, compacting, and upgrading. Also states plainly that each model has its own cache and switching recomputes the entire request even when the content is identical, and that plan mode switches models when the opus plan setting is on. Read 2026-08-21
• Anthropic on deferring tool definitions, and the accuracy result — 2025-11-24. Reports about 77 thousand tokens of startup context coming down to about 8.7 thousand, an eighty five percent reduction, with accuracy improving rather than degrading on their internal evaluation. Treat the accuracy numbers as their claim on their own eval, unreplicated. The post separately notes tool definitions alone reaching 134 thousand tokens before optimisation; that is a different quantity and must not be paired with the 8.7 thousand figure, since 134 to 8.7 is ninety four percent rather than eighty five. In Claude Code deferral became the default in mid January 2026, and it is now unconditional by default rather than threshold based. The ten percent of context window threshold applies only if you opt into threshold mode.
• OpenAI documents deferred tool loading too, which corrects something we said — OpenAI, Tool search guide, read 2026-08-22. States that tool search can be used with deferred functions, namespaces or MCP servers, and that for an MCP server you set defer_loading true on the server tool definition. Codex made this the default rather than a feature flag in pull request 29486, merged 2026-06-22 and shipped around version 0.142.2, which also retired the older condition of a feature flag or at least a hundred tools. We had previously said that OpenAI documented none of this and that the only way to find out was to read the binary. That was wrong. It is in this list because the correction matters more than the claim did
• Manage costs effectively, where the vendor disagrees with itself — Worth reading beside the best practices page. This one says clearing costs nothing, which is true at the moment you clear and says nothing about the cache write on the next call. It also states that a scheduled task fires on its interval while the session is idle, sending your full context each time, and that background usage is typically under four cents a session. Read 2026-08-21
• Claude Code best practices, the other half of the disagreement — Recommends clearing frequently between tasks and clearing after two failed corrections. That second trigger fires around the third to fifth call of a session, which is where our own measurements put setup at roughly seventy nine percent of the input bill, priced at the one hour cache write rate a subscription actually pays. Neither page mentions the other. Read 2026-08-21
• Maximizing the value of your Claude Code sessions — Anthropic, 2026-08-14. The clearest vendor statement of the mechanism this episode is built on: nothing gets sent just once, everything in the conversation gets sent again on every turn after it, and the cache has to match from the very start of the request forward. Also the source for the cache lifetime being an hour on a subscription and five minutes on an API key
• Don't Break the Cache — the academic version of the same finding — Lumer and colleagues, January 2026. Over 500 agent sessions across three vendors. Caching reduces cost by forty one to eighty percent, and naive full context caching can paradoxically increase latency, so they recommend excluding dynamic tool results from the cached prefix. Independent support for the prefix fragility this episode is about
• Effective harnesses for long-running agents — Anthropic, November 2025. Compaction alone is not sufficient for work that spans sessions; durable artifacts such as git history, a progress file and a structured feature list are what actually bridge them. The evidence behind writing down what is true before you clear
• Effective context engineering for AI agents — Anthropic, September 2025. Context as a finite resource, just in time retrieval, structured note taking, and the sub agent return contract. Note the original says sub agents often return summaries of a thousand to two thousand tokens, which describes what happens rather than prescribing a rule
• Context Rot — why a longer window is not a bigger brain — Chroma, July 2025, eighteen frontier models. Every model tested degrades as input grows, and semantically similar distractors hurt more than length alone. Read it for the mechanism rather than for a threshold; the widely repeated three hundred to four hundred thousand token figure is not in this report, which does not test above roughly a hundred and thirteen thousand

The written deep dive — which keeps updating as new work lands — is at AI University.

Sources, every one with its date:

• Building Effective Agents — the five patterns, December 2024 — Anthropic (Erik S. and Barry Zhang)
• How we built our multi-agent research system — the 90.2%, the 15x, and the limitations — Anthropic
• Effective context engineering for AI agents — Anthropic
• Is Graph Engineering Real? Why Everyone Is Talking About It (FOD#159) — Turingpost — Ksenia Se
• Graph Engineering Explained: What Actually Changed — Louis-François Bouchard
• LangGraph — the reference implementation of these patterns since January 2024 — LangChain
• The Ralph Wiggum loop — where the loop rung started — Geoffrey Huntley
• CodeCompass: Navigating the Navigation Paradox in Agentic Code Intelligence — the 58% zero-tool-call finding — arXiv 2602.20048
• Chain-of-thought prompting elicits reasoning in large language models — Wei et al., January 2022
• Retrieval-Augmented Generation for knowledge-intensive NLP tasks — Lewis et al., 2020
• FORGET Loop Engineering. Graph Engineering is about THIS — a worked LangGraph StateGraph — Gao Dalie
• The written deep dive, with the news and videos on this subject updating daily — AI University

The written deep dive — which keeps updating as new work lands — is at AI University.