Sources and evidence

Reading rules#

Five rules were applied to every source. They explain how much weight each number in this research can carry.

Labels are applied to claims that carry a number or could be contested. They are not applied to definitions, to this research's own frameworks (the nine segments, the five constraints, starvation as a failure mode), or to conclusions restated from a labelled claim in the same chapter.

Long-context research#

Chroma, "Context Rot: How Increasing Input Tokens Impacts LLM Performance"#

• Supports: performance generally degrading as input grows across controlled experiments with 18 models, four vendors, 8 input lengths and 11 needle positions; the effect of question–answer similarity; distractors compounding non-uniformly; coherent text retrieving worse than shuffled text across all 18 models; the focused (~300 tokens) versus full (~113K) LongMemEval gap. The curves are non-uniform and model/task-specific.
• Discount: static retrieval tasks, not agent trajectories. Rule R3 applies to every use.
• Used in: chapter 1, chapter 3, chapter 5, chapter 6, chapter 9, chapter 11, chapter 13, lessons A, lessons B.
• Link: trychroma.com/research/context-rot

Effective-length caveat#

Benchmark suites differ in task, model and scoring, so advertised length is not a performance guarantee and Chapter 1 does not compare cross-suite scores.

Liu et al., "Lost in the Middle: How Language Models Use Long Contexts"#

• Supports: a U-shaped position effect in controlled retrieval tasks: relevant information near the beginning or end performed better than information in the middle.
• Discount: the study covers specific models and tasks; the size and shape of the effect vary by model, task and harness. It does not establish a universal utilisation threshold.
• Used in: chapter 1, chapter 13.
• Link: MIT Press / TACL

2026 agent-context papers#

"Toward Reliable Context Compression for Long-Horizon Agents: An Empirical Study of Execution Instability" (TRACE)#

• Supports: seven compression strategies compared on AppWorld (full context, FIFO truncation, token pruning, two prompt-based compaction schemes, guideline-based approaches, and a verifier-guided method); the AppWorld results table (no compression 85.7% / 77.4% Pass²; verifier-guided 77.1% / 67.3%; prompt-based 71.4% / 59.5%; FIFO 63.7% / 53.0%); the Pass@2/Pass² gap widening under tighter budgets; correct termination 44.6% versus 77.2% at 2K; +0.108 marginal POST-minus-PRE increase in blocked/error-action probability at the first post-compaction action in AppWorld.
• Discount: one benchmark (147 API tasks) and specific model pairings; transfer beyond this setting is unverified. The +0.108 result is a marginal effect at that action step, not an absolute count; it identifies a risk at the tested post-compaction boundary. TRACE did not test re-reading a plan file or compare recovery actions.
• Used in: lessons B, chapter 2, chapter 3, chapter 6, chapter 9, chapter 11, chapter 12, chapter 13, chapter 17, case CS-4.
• Link: arXiv 2608.06503

"Addressable Recall Compaction for Long Context-Window Control in AI Agents" (ARC)#

• Supports: the five-baseline comparison on a 1,000-task needle suite and a 311-task LongBench-v2 Hard subset; needle 99.00% / 99.80% versus RAG 79.57% / 96.67%; LongBench-v2 Hard 27.47% / 32.47% versus 25.83% / 30.87%; 38.8–80.3% bandwidth savings; the statement that omitted summary details cannot be recovered.
• Discount: 8B and 32B models, not frontier scale. The reasoning margin is small; do not overstate it.
• Used in: lessons B, chapter 3, chapter 6, chapter 9, chapter 15, chapter 18.
• Link: arXiv 2607.25066

"CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents"#

• Supports: the summariser-only swing, SWE-bench 49.0% → 55.5%; +5.5 / +7.0 on SWE-bench Verified and +6.8 / +3.1 on Terminal-Bench 2.0 from compaction-aware training; operating parameters (10,240-token threshold, at most 3 compactions per run, at most 250 turns).
• Discount: the training results need access most readers lack. The 6.5-point swing at inference time does not.
• Used in: chapter 2, chapter 3, chapter 6, chapter 9, chapter 15, chapter 16, lessons B, case CS-3.
• Link: arXiv 2607.05378

"Self-Compacting Language Model Agents" (SelfCompact)#

• Supports: the authors' rationale that fixed thresholds and intervals can ignore trajectory structure; their task-specific rubric (compact after a sub-task resolves or the trajectory converges; suppress mid-derivation or when stuck); and measured comparisons on competition math and agentic search. Across four Qwen configurations and three math benchmarks, the rubric beat a 16,000-token fixed schedule in 11/12 similarly budgeted cells, with one 1.1-point fixed-interval win. Across three deployed search models and BrowseComp, BrowseComp-Plus and DeepSearchQA, it had higher accuracy in all nine cells than the 30%-of-context threshold baseline. Figure 1 is one BrowseComp trace where a fixed interval drops four verified facts; it is an illustration, not an aggregate result. Other reported findings include summarisation as a blocking call that can stall the agent for tens of seconds and prompt instructions about summary length being largely ignored.
• Discount: the broad trigger critique is author motivation, not a general experiment. Results apply to these models, benchmarks, budgets and this rubric; they do not show that all threshold or periodic triggers fail or that the rubric transfers to coding-agent workflows. The rubric is one instantiation.
• Used in: chapter 3, chapter 6.
• Link: arXiv 2606.23525

"The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management"#

• Supports: masking matching LLM summarisation on SWE-bench solve rates across two model sizes at substantially lower cost; preferring simple methods first.
• Discount: the numeric detail was captured from the paper's summary rather than a full read. Re-verify before citing externally.
• Used in: chapter 6, chapter 9, chapter 11, chapter 13, chapter 18.
• Link: arXiv 2508.21433

"Is Grep All You Need? How Agent Harnesses Reshape Agentic Search"#

• Supports: grep generally beating vector retrieval on 116 LongMemEval-derived questions; four harnesses compared, inline versus file-based results; overall scores depending strongly on harness and tool-calling style on identical data; eight surveyed agents using the model as a navigator over shell tools.
• Discount: the questions are conversational-memory shaped, not repository shaped. The harness finding transfers more confidently than the grep-versus-vector ranking.
• Used in: lessons B, chapter 3, chapter 5, chapter 9, chapter 11, chapter 13, chapter 18, case CS-7.
• Link: arXiv 2605.15184

Token economics of tool-heavy agents (including arXiv 2606.10209)#

• Supports: across four GPT-5 configurations averaged over five runs on a 50-task Dynamics 365 hotel-expense benchmark using verbose MCP responses, input tokens were 99.75–99.87% of total token volume. Full context used 2.68× the total tokens of the best managed configuration and completed fewer tasks.
• Discount: one structured form-filling workflow with verbose accumulated tool responses; not representative of agents generally. The input-token share is not a cost share: generated output has separate model/provider rates, while input cost depends on cache status and provider pricing. Measure segment volume, task outcomes and priced cost on the target workload.
• Used in: chapter 2, chapter 3, chapter 5, chapter 9, chapter 10, chapter 11.
• Link: arXiv 2606.10209

Practitioner engineering writing#

Anthropic, "Effective context engineering for AI agents"#

• Supports: just-in-time retrieval with lightweight identifiers; progressive disclosure; metadata as a relevance signal; compaction keeping decisions and unresolved bugs while discarding redundant output.
• Discount: first-party, describing its own products. Read for mechanism.
• Used in: chapter 3.

Anthropic, multi-agent research system write-up#

• Supports: the orchestrator-worker architecture; 90.2% over single-agent Opus 4 on an internal research evaluation; about 15× chat tokens (agents about 4×); about 80% of performance variance explained by token usage.
• Discount: research tasks, not coding; an internal evaluation; first-party. The 15× and 80% must always be quoted with the 90.2%.
• Used in: chapter 3, chapter 7, chapter 9, chapter 11, chapter 13, lessons B, case CS-2.

Sub-agent token-efficiency reporting#

• Supports: about 9K total tokens for a multi-domain query answered through isolated sub-agents versus about 15K with an accumulating-context pattern.
• Discount: P. A single practitioner comparison with no published method; no independent reproduction found.
• Used in: chapter 3, chapter 7, chapter 11.

Cognition, "Don't Build Multi-Agents"#

• Supports: share full traces, not messages; actions carry implicit decisions, which conflict when parallel agents build interlocking parts; the Flappy Bird failure; single-threaded linear agents.
• Discount: first-party, arguing for its own architecture. The failure description is the durable content.
• Used in: chapter 4, chapter 7, chapter 12, case CS-1.

LangChain, write / select / compress / isolate, and Deep Agents#

• Supports: the four-operation taxonomy; filesystem offload; compression middleware.
• Discount: framework-specific implementation details.
• Used in: chapter 1.

Sourcegraph, context-engineering guidance#

• Supports: the author reports that 5K-token targeted retrieval outperformed a 100K-token codebase summary on the same task P. This is an author-described practitioner observation, not a published benchmark result.
• Discount: the first-party, vendor-authored post says “we've seen” but identifies no task, model, harness, prompt, comparison protocol, trial count or results table. Treat it as a bounded observation and a hypothesis to test locally; it does not establish a general advantage for smaller inputs or a universal token-size rule.
• Link: Sourcegraph, “Context Engineering: A Practical Guide for AI Agents”.
• Used in: chapter 3, chapter 5, chapter 11, chapter 13, lessons B.

Hybrid code-retrieval reporting#

• Supports: a hybrid of semantic search and grep reported 12.5% more accurate than either retrieval strategy alone.
• Discount: P. No published method and no independent reproduction found.
• Used in: chapter 3, chapter 5, chapter 11, chapter 13.

Practitioner search-time reporting#

• Supports: agents spending 60%+ of their time — and similarly of their turns — locating context rather than editing.
• Discount: P. Single-practitioner reports with no published method; the time-based and turn-based denominators were reported separately.
• Used in: chapter 5.

Drew Breunig, "How Long Contexts Fail"#

• Supports: the poisoning, distraction, confusion and clash taxonomy, a commonly used taxonomy rather than a measurement.
• Discount: a taxonomy, not a measurement. This research adds starvation as a fifth mode.
• Used in: chapter 1.
• Link: dbreunig.com

Tool-pool size and tool selection#

• Supports: RAG-MCP reports 43.13% accuracy versus 13.62% for blank conditioning in a method comparison on the held-out web-search subset of MCPBench; it does not report those numbers as a decline caused by increasing tool count. In a separate stress test, the paper varies a candidate MCP pool from 1 to 11,100 over 20 web-search tasks, with one relevant MCP and distractors; it reports a non-monotonic pattern and marked degradation as pools become large. Speakeasy's Dog API demonstration reports 19/20 successful tool calls for Qwen3 1.7B at 20 tools, 3/4 at 40, and frequent hallucinations/errors at 107 P. The same post says its demonstration is not exhaustive or rigorous.
• Discount: RAG-MCP is a paper-specific evaluation; its method comparison and increasing-pool stress test are distinct experiments and do not establish a universal tool-count threshold. The Dog API result is a vendor/practitioner demonstration with a synthetic one-tool-per-breed setup, a small model, and very limited trials; its 107-tool observations are not “complete failure” across models. No cited result supports fixed 20-tool or 40-tool ceilings. The ~42,000-token server report and quarterly regrowth observation remain practitioner claims P.
• Used in: chapter 2, chapter 3, chapter 8, chapter 9, chapter 13, chapter 15, lessons B.
• Links: RAG-MCP; Speakeasy, “Why less is more for MCP”.

Progressive disclosure and code-execution reporting#

• Supports: ~25,000 → ~2,500 tokens for descriptions versus definitions; 150,000 → ~2,000 (98.7%) for code execution; 99%+ on definitions at 112 tools; an independent 98% production report on a GitHub MCP server; the practitioner-landscape convergence on deferred definitions, with some harnesses making them the default by 2026, and usage reporting / compaction control increasingly built in; emerging MCP proposals for adaptive response granularity.
• Discount: first-party and community reports; no independent reproduction found.
• Used in: lessons B, chapter 3, chapter 8, chapter 15, chapter 17, chapter 18, case CS-6.

Prefix-caching operational reporting#

• Supports: 85.2% hit rate with ~46,059 tokens reused per request; 90% hit rate giving sub-200 ms time to first token and 80–90% compute savings; byte-exact prefix matching.
• Discount: P, specific to particular serving stacks. The relative prices in chapter 10 are generic shapes, not any provider's rate card.
• Used in: chapter 1, chapter 3, chapter 10, case CS-5.

AGENTS.md and instruction-file practice#

• Supports: 30+ tools reading it; 60,000+ repositories; Agentic AI Foundation stewardship; the ≤150-line guidance; ~4% success improvement from human-written context files.
• Discount: P throughout. The 4% figure has no published method.
• Used in: chapter 2, chapter 5, chapter 11, chapter 13, chapter 17, chapter 18.

Open-source tooling documentation#

• Supports: the tooling catalogue; the 60–90% and 98% output-reduction claims; symbol-level operations; language coverage (30+ languages via the Language Server Protocol across the surveyed toolkits).
• Discount: self-reported, none reproduced here. Audit before installing.
• Used in: chapter 5, chapter 10, chapter 16, chapter 17, chapter 18.

Verification caveat#

Primary sources were read directly for the load-bearing claims: the compaction cluster (TRACE, ARC, CompactionRL, SelfCompact), the masking comparison, the agentic-search harness study and the context-rot study. Several supporting figures were captured from abstracts, search summaries or secondary coverage instead of a full paper or published protocol.

Figures in that category:

• the observation-masking numeric detail
• the ~42,000-token single-server figure
• the Speakeasy Dog API tool-count demonstration; its setup and limited scope are now stated above
• the hybrid +12.5% figure
• the 9K-versus-15K sub-agent token comparison and the within-a-quarter regrowth observation
• the 60%+ search-time figure
• the progressive-disclosure family of figures (25K→2.5K, 150K→2K / 98.7%, 99%+ at 112 tools, the 98% GitHub report)
• the AGENTS.md adoption and ~4% figures
• the 85.2% / 46,059-token prefix-cache figures
• every percentage in the tooling chapter

Re-verify any of these before citing them externally. Every pre-existing figure in this list is corroborated in direction by at least one S source and none is the only support for a load-bearing conclusion. The three figures added during the chapter 3 review — the hybrid +12.5% result, the 9K-versus-15K sub-agent comparison and the within-a-quarter regrowth observation — stand on single practitioner reports and are treated as upper bounds wherever they appear.

What would prove this research wrong#

Six observations would materially undermine the conclusions. Each is cheap to test on your own workload, which is the point of stating them.

1. A frontier model with flat performance to about 80% of its window on a distractor-rich task. That would undercut the attention-budget framing and much of the subtraction default.
2. A controlled study where Pass² degrades no faster than mean accuracy under compression. That would remove the variance-first finding and make single-run evaluation legitimate again.
3. A retrieval ranking that replicates across three independent harnesses with similar margins. That would weaken the harness-confound rule and make published comparisons transferable.
4. A replicated, representative tool-selection evaluation finding no meaningful change in task success or wrong-tool rate as the candidate pool grows, while controlling for prompt size and harness. That would weaken the claim that larger choice sets can create a selection burden.
5. Parallel sub-agents reliably producing compatible interlocking implementations without a shared brief. That would invalidate the composability test.
6. Total token spend falling when delegation is introduced. That would change the isolation economics entirely.