Frequently asked questions

Because performance can become less reliable as input grows, even before the window is full. A study of 18 models from four vendors/model families found the direction across controlled tasks, with model- and task-specific variation S. A bigger window gives you more room to fit things, not a guarantee that the model will use them well. See chapter 1.

The everyday name for performance becoming less reliable as input length grows. Chroma measured it across 18 models; the size and shape of the effect varied by model and task. It also worsened with vocabulary gaps, near-duplicate distractors and coherent surrounding prose. See chapter 1.

Position, not memory, can matter. In cited retrieval tasks, relevant information in the middle performed worse than information near the beginning or end; the effect depends on model, task and harness S. Keep critical rules near the action and verify late-session behavior. See position decay.

Four numbers, in about 30 minutes: your prefix tax (send . in a fresh session and read the input count), tool-definition tokens versus tools actually called, tool output as a share of the session, and cache hit rate. See Phase 0.

There is no well-supported universal cutoff. One RAG-MCP paper reports 43.13% accuracy for retrieval-based selection versus 13.62% for blank conditioning; its separate increasing-pool stress test uses one relevant server plus distractors and reports non-monotonic degradation at large pool sizes S. A vendor's small Dog API demonstration reports 19/20 calls correct for Qwen3 1.7B at 20 tools, 3/4 at 40, and frequent errors at 107; the authors say it is not exhaustive or rigorous P. Measure wrong-tool rate, task success, prompt cost, and missed capabilities on your own representative tasks before setting a local budget. See chapter 8.

Only half of it. Deferred definitions fix displacement (tokens crowding out work). They do not fix selection confusion, because the agent still chooses among the same number of options. Only fewer tools, or per-task scoping, fixes confusion. See chapter 8.

Prefer the shell for anything it already does: gh, psql, curl, rg, git. It composes with pipes, loads no schema, and is well represented in training data. Keep an MCP server deliberately when it is a security control, and write that decision down. See chapter 8.

Probably the cache. Prompt caching is prefix-exact: changing anything early in the prompt invalidates everything after it. A worked example cut tokens 24% by loading tools dynamically and raised its input-cost estimate 6.9× C. Measure cached and uncached input at provider rates, then add generated output at its separate model/provider rate; raw token volume is not total cost. See chapter 10.

Not bad, but costlier than it looks. Changing only the summariser moved SWE-bench 6.5 points, and compression made agents intermittent before it lowered their average S. Compact at a sub-goal boundary, at most once per session, after offloading large outputs, and with an explicit schema. See chapter 6.

If you keep a plan file, start a new session. A reset costs about 500 tokens to reload the plan, removes every accumulated distractor and poisoned fact, and lets you author exactly what crosses the boundary. A second compaction in one session is the signal to reset. See chapter 3.

Goal; state as done, in progress and not started; decisions with reasons; ruled-out approaches with evidence; exact strings verbatim (errors, versions, paths, line numbers); open questions; offloaded artifacts; and anything unverified marked as such. See the compaction schema.

Usually not. The repository is the best memory: ADRs, tests and comments are versioned, reviewed and repaired with the code. Use a memory tool only for expensive-to-learn facts with no repository home, and delete it if it produces fewer than one retrieval a week that changes an action. See chapter 6.

Let it fetch a page by name when a person points to it, and treat the page as a hypothesis to check against the code. Do not bulk-index the wiki: stale pages about exactly your module are the most plausible distractors available. Migrate durable facts into the repository instead. See chapter 6.

Grep and the symbol graph first. In a four-harness comparison grep generally beat vector retrieval, and the harness mattered more than the strategy S. Use semantic search as a complement when you do not know the name of what you are looking for, and point it at docs and ADRs. See chapter 5.

Give it pointers, not prose: entry points, invariants, landmines and commands, under 2,000 tokens. A Sourcegraph author reports 5K-token targeted retrieval outperforming a 100K-token codebase summary on the same task P (see Sourcegraph's post), but the post names no task, model or comparison protocol; treat this as a vendor/practitioner observation, not benchmark evidence. Separately, coherent documents retrieved worse than fragments in the cited tests S. See chapter 5.

For read-mostly work whose outputs simply add up: finding callers, surveys, test triage, dependency audits. Not for interlocking implementation, and never for debugging. If you cannot write the sub-agent's output schema in 20 lines, do not delegate. See chapter 7.

Each sub-agent pays its own prefix tax, so parent context falls while total spend rises. A published research system used about 15× the tokens of a chat, and spend alone explained about 80% of its gain P. See chapter 7.

Run the same tasks through both configurations, at least twice each, and compare Pass² (solved on both runs) with McNemar's test. On 50 tasks, 68% versus 74% is not a result: the intervals overlap almost entirely. See chapter 9.

Pass² counts a task as solved only if it succeeds on both of two runs; Pass@2 counts it if either succeeds. Compression damages Pass² about three times more than Pass@2, so it is the metric that catches "it worked yesterday". See chapter 9.

When cache hit rate is over 70%, fewer than 3 tools are defined per tool used, tool output is under 35% of the session, read-utilisation is over 15%, post-compaction re-fetches are under 2, and Pass² ÷ Pass@2 is over 0.85. Past that, your constraint is elsewhere. See the stop rule.