⌖ AI Field Guide

Anthropic has deprecated this on Claude 4.6 models and removed it entirely on Claude 4.7+, where requests using type: 'enabled' now return a 400 error, in favor of adaptive thinking plus the effort parameter. Still required on models that only support extended-thinking mode (Opus 4.5, Haiku 4.5, Sonnet 4.5 and earlier), so not dead everywhere, but no longer the recommended path for new work.

Anthropic's own docs mark the manual budget_tokens mechanism as deprecated on Claude 4.6 models and rejected outright (400 error) on Claude 4.7 and later -- adaptive thinking is the only mode going forward wherever it's available. Adaptive mode can also skip thinking entirely on easy inputs at lower effort levels, which a fixed budget can't do.

This isn't theoretical: researchers have disclosed a steady cadence of MCP CVEs through 2026, including real prompt-injection bugs in an official Anthropic-maintained MCP server, and most organizations experimenting with MCP haven't put it through formal security review before reaching production. OWASP's MCP cheat sheet mitigations (least privilege, definition pinning, sandboxing, human-in-the-loop, treating tool output as untrusted) match what the incident writeups say would have helped.

Anthropic's own 'demystifying evals' guidance lays this out as field-tested methodology: early on, effect sizes are large so small samples suffice, and the bar shifts to harder evals and tighter calibration only once an agent starts saturating existing tests. The discipline matches current LLM-as-judge practitioner writing too -- judges carry position/length/self-preference bias and need the same calibration rigor as any other measurement tool.

Sources genuinely disagree. Cost/latency data favors RAG heavily (cited near $0.00008/query vs. $0.10-6.75 for 100k-1M-token long-context calls, responding in under 2s vs. 30-60s); long-context proponents point to it beating chunk-based RAG on QA quality when the source is stable and fits the window. 'RAG is dead' and 'in defense of RAG' pieces are both circulating in 2026 with real evidence behind each -- the right call depends on corpus size/stability and latency/cost budget, not a universal winner.

Anthropic reports this combination cutting top-20-chunk retrieval failure from 5.7% to 2.9% (49% fewer failures), and to 1.9% (67% fewer) with reranking added -- each layer contributes independently. It remains the reference approach for chunk-based RAG rather than having been superseded by something else.

Anthropic reports this winning specifically for parallelizable research, context isolation, and tool-set specialization -- their research system beat single-agent Claude by over 90% on their internal eval. But it's not a default: multi-agent setups use 3-10x the tokens of a single agent, and uncoordinated multi-agent setups have been shown to amplify errors far more than centrally-orchestrated ones. Anthropic's own advice is to start with one well-tooled agent and add subagents only once that proves insufficient.

Anthropic's tool-design guide frames tools as a contract between deterministic software and a non-deterministic reasoner, and reports that small wording and consolidation changes produced outsized gains on real evals -- consistent with their broader finding that ambiguity in a tool list becomes ambiguity in agent behavior.

Demonstrated directly in Anthropic's 'Claude plays Pokemon' example, which tracked precise progress (e.g. exact level-ups) across thousands of steps via a notes file rather than in-context state, and recommended in Anthropic's context-engineering guidance as a standard long-horizon technique -- cheap, auditable, and it survives compaction and restarts in a way in-context state doesn't.

Anthropic lists compaction alongside just-in-time retrieval and note-taking as a standard technique for running agents past a single context window; clearing stale tool results is called out as especially low-risk since raw outputs are usually only needed briefly after the call. Caveat: recent research on a 'compaction cliff' finds naive summarization degrades sharply past a point, so what's kept vs. dropped should be deliberate, not automatic.

Anthropic's 2026 context-engineering guidance names this as a core technique for long-horizon agents, alongside compaction and note-taking: preloading everything burns context budget on data that's never used and crowds out tokens the model needs for actual reasoning, mirroring how people use indexes/filesystems instead of memorizing whole corpora.

Keeps the model's working context focused on the task at hand, lets instructions be authored and updated independently, and scales to far more accumulated knowledge than a single file can hold.

A single sprawling context file crowds out the model's working context, goes stale as the project changes, and can't be selectively loaded -- Anthropic's own engineering guidance now recommends small, scoped files/skills loaded progressively instead.