Method, scope and quality bar

The research question#

At inference time, a model can act on two sources: patterns encoded in its weights and information placed in the current input. The repository, yesterday's conversation, CI logs and the ticket affect the next action only when the harness retrieves and presents them. That machinery is rarely designed as one system. It accretes.

The question is not "how do I fit more into the window?" That has an answer (buy a bigger window), and the answer does not work. It is:

The strong claim under test: before changing the model or adding tools, teams running long repository tasks should test whether context discipline is their binding constraint.

Context failures are also the quietest. An agent with a bad loop spins visibly. An agent with bad evaluation reports a number you can argue about. An agent with bad context management quietly gets things wrong, and the transcript looks reasonable the whole way down.

Sub-questions and where they are answered#

Eight design commitments#

Fixed before writing, to avoid the usual failure of "context engineering best practices" content: mostly restated blog posts, some vendor pitch, very little measurement.

Method#

1. Ground truth. Three bodies of work read against each other: empirical long-context research (for the shape of the degradation curve), the 2026 agent-context papers (for measured changes at compaction boundaries) and practitioner engineering writing (for mechanism, discounted for product claims). Vendor numbers about vendor products are upper bounds.
2. Derive the frame. A segment entered the anatomy only if a documented class of failure is exactly "this segment was too large, too small, stale or in the wrong place".
3. Catalogue. Ten methods chosen by four tests (mechanism, evidence, independence, decidability), with a published benched list. See how the ten were chosen.
4. Measurement layer. Evaluation design specific to context changes, because their effects are small, fully harness-confounded, and show up in variance first.
5. Apparatus. Hard calls, failures, metrics, antipatterns, cases, tooling, lessons, habits, templates and a scored diagnostic.
6. Personalise. A sequenced plan with a measurement gate per phase, so nothing is adopted on faith.

Out of scope#

Stated so that the omissions are choices, not oversights.

• Model internals: attention variants, kernel-level cache eviction, positional encodings. Referenced only where they explain observable behaviour.
• Training-time context: long-context pretraining and compaction-aware training appear only where they change what an operator should do today.
• Non-coding agents: browser, GUI and research agents appear only where a finding transfers. Coding agents are special because the filesystem is simultaneously the task, persistent memory and a major evidence source; it is not complete ground truth.
• Prompt quality: how to write a good instruction is prompt engineering. How much of it to load, when and at what cost is context engineering.
• Security: context poisoning through prompt injection is named as a failure with containment; the adversarial surface is another research programme's subject.

The acceptance bar#

The work was not considered done until 27 criteria passed, plus two "anti-bar" conditions that fail the work even if every criterion passes. Result: 27 of 27 passed. Three passed with stated caveats. Both anti-bar conditions passed.

The three caveats#

• Case studies, "majority sourced", is true only on a generous reading. Of eight cases, four are S, two are P first-party accounts, and two are labelled composites C. Strictly, S is a plurality, not a majority. Six of eight describe real events; four have a published method.
• Derived claims are under-labelled. Rule R3 (long-context findings transfer to agents only by inference) applies far more widely than the explicit D labels suggest. Treat every application of static-text research to a live agent as derived, whether or not it carries the label.
• One of the central empirical claims is under-verified. The observation-masking result was captured from the paper's summary. It is corroborated in direction and is not the sole support for any conclusion, but re-read it before citing. The sources page lists every figure in this category.

The two anti-bar conditions#

1. If a reader with a real agent cannot, after reading the methods and the diagnostic, name their single largest wasteful segment and the change that would shrink it, the work has failed. It passes: the one-question version branches to a named action, and a reader whose largest segment is tool definitions reaches the representative-sample deletion check in under fifteen minutes.
2. If any section could be replaced by "just use a bigger window" or "just run /compact" without loss, it is filler. It passes, and the research argues the opposite of both: more context is not a reliability guarantee, and naive compaction is where the damage is.

Relationship to sibling research#

About this edition#

This web edition was rewritten from the original research package (about 72,000 words across 32 files, locked 2026-09-05) for readability and navigation: plain language, one question per chapter, a learning path in five parts, figures and charts, a one-page summary, an FAQ and a glossary. Numbers, evidence labels, caveats and falsifiers are carried over unchanged. Where the original used internal section references, this edition uses links.

How to cite#

Alhamoud, G. (2026). Context Management in Coding Agents (RSCH-001, v1.0).
https://ghassan-alhamoud.com/research/context-management-coding-agents/