Glossary

A

Addressable recall

Compaction that replaces old content with content-addressed stubs the agent can fetch on demand, instead of a paraphrase. Lossless by construction.

Measured at +19.43 points over the best lossy baseline on a 1,000-task needle benchmark S.

Also: addressable recall compaction, ARC

Introduced in Chapter 6: Compaction, offload and memory

ADR

A short, numbered Markdown record of a design decision, committed to the repository and reviewed in the pull request that makes the decision.

The best memory for coding agents in this research: versioned, greppable and repaired by the same process that changes the code.

Also: architecture decision record, ADRs

Introduced in Chapter 6: Compaction, offload and memory

Agentic search

A retrieval control strategy where the agent iterates: search, read, decide the next query, stop when satisfied.

The dominant architecture in shipped coding agents, and the axis where harness differences bite hardest.

Also: agentic retrieval, live agentic search

Introduced in Chapter 5: Retrieval: how code should enter context

Attention budget

The framing that replaces "context window as container": adding tokens can redistribute useful signal rather than creating more reliable attention.

Introduced in Chapter 1: Why more context can hurt coding agents

B

Binding constraint

The one limitation that currently caps your agent's results. Fixing anything else changes little until it moves.

Introduced in Chapter 4: Choosing your methods: a 55-minute procedure

Body

Segments 5–9 of the context: retrieved code, tool results, agent messages, user turns and summaries. It grows during a session.

Appended content may be reusable once unchanged; cache hits and billing depend on the provider and request construction. Contrast with the prefix.

Also: context body

Introduced in Chapter 2: Anatomy of a coding agent's context

C

Cache-adjusted cost

Input cost with cached and uncached tokens priced separately, about 0.10 versus 1.25 in relative units.

The only honest cost metric, because context changes often move the ratio rather than the total.

Introduced in Chapter 10: Metrics and the real cost of context

Clash

A failure mode where the context contains contradictions from different phases and the model resolves them by recency or by blending.

Introduced in Chapter 1: Why more context can hurt coding agents

Code graph

The import, call, reference and type graph of a codebase, computed exactly by a compiler or language server.

Free, always current, and the basis of structural retrieval. Not the same thing as a knowledge graph.

Also: reference graph, call graph

Introduced in Chapter 5: Retrieval: how code should enter context

Compaction

Replacing part of a transcript with a summary. Irreversible, within one session, and it invalidates the cache from the point of change.

Distinct from offload (reversible) and memory (across sessions).

Introduced in Chapter 6: Compaction, offload and memory

Composability

The property that makes delegation safe: sub-agent outputs combine without negotiation.

Search and survey work is composable. Interlocking implementation is not.

Introduced in Chapter 7: Sub-agents and isolation

Confusion

A failure mode where irrelevant context is used because it is present. The common case is picking a plausible wrong tool when too many are in scope.

Also: tool confusion

Introduced in Chapter 1: Why more context can hurt coding agents

Context rot

The everyday name for performance becoming less reliable as input length grows. Chroma measured the pattern across 18 models, with non-uniform curves that depend on model and task [S].

Introduced in Chapter 1: Why more context can hurt coding agents

Contract loss

A sub-agent knew something relevant but did not report it because its output contract did not ask.

The parent cannot recover it unless the sub-agent's transcript was saved.

Introduced in Chapter 7: Sub-agents and isolation

D

Dereference, don't index

The policy for external prose such as a wiki: fetch a named page on demand, never bulk-ingest it into a retrieval store.

Also: dereference

Introduced in Chapter 6: Compaction, offload and memory

Distraction

A failure mode where the agent imitates its own transcript instead of reasoning: re-running, re-reading, re-proposing.

Introduced in Chapter 1: Why more context can hurt coding agents

E

Effective context length

The length at which a model still performs acceptably on a given task, as opposed to the length it advertises; it may be shorter than the advertised maximum.

Introduced in Chapter 1: Why more context can hurt coding agents

Evidence label

A tag on a claim saying how it is supported: S (sourced), P (practitioner-reported), D (derived) or C (composite).

Also: evidence labels

Introduced in Chapter 1: Why more context can hurt coding agents

F

Fragment blindness

Reasoning about a code fragment without the control flow around it, such as a grep hit shown without the if statement that governs it.

Introduced in Chapter 5: Retrieval: how code should enter context

H

Harness

The software wrapping the model: system prompt, tools, loop, compaction and retrieval behaviour.

Harness effects can exceed the effect of the strategy under test, so results rarely transfer between harnesses S.

Also: agent harness, harness confound

Introduced in Chapter 2: Anatomy of a coding agent's context

I

Isolation

Splitting context across boundaries such as sub-agents or sessions. It shrinks the parent's context at the expense of total spend.

Introduced in Chapter 7: Sub-agents and isolation

J

Just-in-time retrieval

Carrying lightweight identifiers such as paths and symbol names, and fetching content only at the moment it is needed.

Also: JIT retrieval, just-in-time

Introduced in Chapter 3: Ten methods that survive scrutiny

K

Knowledge graph

Entities and relations extracted from prose by an LLM pipeline. Approximate and as stale as its last ingestion.

Skipped for coding work: over source code the code graph answers the same questions exactly. Useful for cross-system relationships.

Introduced in Chapter 6: Compaction, offload and memory

L

Laundering

Compaction turning a hedged hypothesis into an asserted fact by stripping the uncertainty around it.

Also: hallucination laundering

Introduced in Chapter 1: Why more context can hurt coding agents

Loop detector

A check that flags the same tool called with the same arguments three or more times with no new information in between.

The cheapest high-signal instrument available, and rarely built.

Introduced in Chapter 1: Why more context can hurt coding agents

M

Masking

Replacing old tool output with placeholders. Needs no model call.

Matched LLM summarisation on solve rate at lower cost S, so it is the baseline any summariser must beat.

Also: observation masking

Introduced in Chapter 6: Compaction, offload and memory

MCP

The Model Context Protocol: a standard way to expose tools and data to agents through servers.

The protocol is fine. Attaching many servers by default is the problem.

Also: Model Context Protocol, MCP server, MCP servers

Introduced in Chapter 8: The tool surface: MCP, skills and progressive disclosure

Memory

Information persisted across sessions. Justified only by retrievals that changed an action, not by what was written.

Also: cross-session memory

Introduced in Chapter 6: Compaction, offload and memory

O

Offload

Moving content out of the window into addressable storage and leaving a stub the agent can follow back. Reversible.

Also: reversible offload

Introduced in Chapter 3: Ten methods that survive scrutiny

P

Pass^k

Pass@k counts a task solved if any of k runs succeed ("can it ever?"). Pass^k, written Pass² for k = 2, requires all k runs to succeed ("can it reliably?").

Compression damages Pass^k far more than Pass@k, which is why single runs miss it.

Also: Pass², Pass@k, pass at k, pass-hat-k, Pass@2

Introduced in Chapter 9: Evaluating a context change

Poisoning

A failure mode where a false fact enters the context and is afterwards treated as established. It survives compaction and cannot be removed by appending a correction.

Introduced in Chapter 1: Why more context can hurt coding agents

Position decay

A critical fact becoming harder to use because its position in a long input is less favourable.

The size and shape of the effect depend on the model, task and harness S.

Introduced in Chapter 1: Why more context can hurt coding agents

Prefix

Segments 1–4 of the context: system prompt, tool definitions, instruction files and skill descriptions. Usually stable and sent on each call; it may be reusable when unchanged, while billing depends on cached versus uncached rates.

Also: prompt prefix, stable prefix

Introduced in Chapter 2: Anatomy of a coding agent's context

Prefix caching

Reusing the computed state of an unchanged prompt prefix. Prefix-exact: valid up to the first changed byte.

Also: KV cache, prompt caching, cache

Introduced in Chapter 1: Why more context can hurt coding agents

Prefix tax

The tokens present in segments 1–4 on each call before any work. Measure them by sending a one-token message in a fresh session; billing depends on cache hits.

Introduced in Chapter 2: Anatomy of a coding agent's context

Progressive disclosure

Loading short descriptions first and full content only on demand. Applied to tool schemas, skills and file reads.

About 25,000 tokens of tool definitions became about 2,500 tokens of descriptions S.

Also: deferred tool definitions, deferred definitions

Introduced in Chapter 8: The tool surface: MCP, skills and progressive disclosure

R

Read-coverage

Whether the decisive file was ever opened during a failed task. It separates starvation from dilution, which need opposite fixes.

Introduced in Chapter 5: Retrieval: how code should enter context

Read-utilisation

The share of retrieved tokens that appear in the final change or explanation. Low values mean retrieval is wasteful.

Introduced in Chapter 2: Anatomy of a coding agent's context

Relevance density

Tokens that could plausibly be cited in a correct answer, divided by all tokens in context. The working measure of dilution.

Introduced in Chapter 1: Why more context can hurt coding agents

Reset

Ending a session and starting a fresh one seeded with a plan file or handoff note. The only operation that removes poisoned content.

Also: session reset

Introduced in Chapter 3: Ten methods that survive scrutiny

Retention integral

The sum of each segment's tokens multiplied by the number of turns it stays in context. The true cost of hoarding.

Introduced in Chapter 2: Anatomy of a coding agent's context

S

Semantic triggering

Compacting on task events such as a sub-goal closing or a test passing, instead of at a token threshold or on a timer.

Introduced in Chapter 6: Compaction, offload and memory

Skill

A named capability with a short trigger description that is always loaded and a full procedure that loads only when it matches.

Also: skills, agent skills

Introduced in Chapter 8: The tool surface: MCP, skills and progressive disclosure

Staleness rate

How fast a fact stops being true. Where a fact should live depends on whether the store's repair rate beats it.

Introduced in Chapter 6: Compaction, offload and memory

Starvation

The agent lacks a fact, does not know it is missing, and proceeds on an assumption. A clean trajectory and a wrong answer.

The fifth failure mode this research adds. It can survive review because the transcript may look clean.

Introduced in Chapter 1: Why more context can hurt coding agents

Structural retrieval

Retrieving code by symbol, reference and import graph instead of by whole file or embedding similarity.

Also: symbol-level retrieval

Introduced in Chapter 3: Ten methods that survive scrutiny

Stub

What stays in context after an offload: an identifier plus just enough description to decide whether to fetch it.

A bare identifier is dead weight; a pointer the agent will not follow is equivalent to deletion.

Also: stubs, citation stub

Introduced in Chapter 6: Compaction, offload and memory

Sub-agent

A separate agent with its own fresh context, given a bounded task and returning a result to a parent agent.

Also: sub-agents, subagent

Introduced in Chapter 7: Sub-agents and isolation

T

Tool bloat

Tool definitions taking a large share of the prefix. Harms through displacement and through selection confusion.

Introduced in Chapter 8: The tool surface: MCP, skills and progressive disclosure

W

Wilson interval

A confidence interval for a success rate that behaves well at small sample sizes.

Also: Wilson score interval

Introduced in Chapter 9: Evaluating a context change