Fundamentals glossary

The reference companion to the fundamentals series (the layers · context · harness · the gate · orchestration): every term those pages use, alphabetized, dated where a date exists, and graded by the same rubric — so you can look one up mid-sentence instead of re-reading five pages. Write "—" where a date or attribution is genuinely unknown; nothing here is a guess dressed up as a fact.

Term	Named	By	What it means	Grade
Ablation	—	—	Remove one component of a system and re-measure to isolate its causal effect; borrowed from ML/statistics, not an agent-engineering coinage	E3
Agentic loop	30 Sep 2025	Simon Willison ("Designing agentic loops")	The repeated call-model-then-act-then-feed-result-back cycle inside one agent run; "an agent runs tools in a loop to achieve a goal"	E2
Backpressure	14 Jul 2025	Geoffrey Huntley ("The Ralph technique")	Verification, not generation, is the scarce resource once an agent can produce a plausible fix in seconds — the honest name for what a gate is for	E2
Compaction	—	—	Summarizing or discarding older conversation/tool-output history so a session stays inside its context budget; ships as a runtime feature	E3
Context engineering	18 Jun 2025	Lütke coined; Karpathy amplified (25 Jun 2025); Anthropic gave the canonical definition (29 Sep 2025)	Curating the smallest set of high-signal tokens the model sees, not maximizing what fills the window	E2
Context rot	—	Chroma (research finding)	Measured accuracy degradation as input length grows, across 18 tested frontier models, even with irrelevant filler tokens	E1
Evaluator-optimizer	19 Dec 2024	Anthropic ("Building Effective Agents")	One model drafts, a second critiques in a loop, with explicit stopping conditions and ground truth from the environment	E2
Fan-out	—	—	Splitting one piece of work across several concurrent agents that each handle an independent slice, then reconciling their results	E3
Gate	—	—	The part of a loop that can say no — an executing test, a fresh-context reviewer, a spend cap, a human at the merge button	E3
Generator/evaluator split	—	—	Structural separation of the agent that produces a change from the agent that judges it, on a fresh context, because the author can't see its own chain of self-persuasion	E3
Graph engineering	18 Jul 2026	A satirical tweet (Peter Steinberger)	Treating agent orchestration as graph-shaped rather than loop-shaped; carries at least 3-4 incompatible meanings	E4 (discipline); E1-E2 (older, unrelated knowledge-graph/GraphRAG research now conflated with it)
Harness engineering	5 Feb 2026	Mitchell Hashimoto named; Addy Osmani popularized (19 Apr 2026)	Everything around the model that isn't the model — tool surface, permissions, sandboxing, the loop, durable state — and the evidence says it moves outcomes more than the model does	E2
Loop engineering	Jun 2026	Addy Osmani named; Simon Willison and Geoffrey Huntley practiced it unnamed a year earlier	The unit of work is "generate, check, critique, fix, repeat, until a gate says done" — a loop is what makes a harness repeat itself	E3
Maker≠checker	—	—	The agent that wrote a change must not be the one that judges it; a different context (ideally a different model) evaluates instead	E3
Prompt engineering	~2023	—	The wording of one request to the model; scoped down from "the whole job" once context, harness, and loop absorbed the rest	E3
ReAct	Oct 2022	Yao et al. (arXiv:2210.03629)	Reason, act, observe, repeat — the think-act-observe cycle a single agent run executes; now described as the loop living inside the harness	E2
Subagent	—	—	An isolated agent instance a driver delegates to, with its own context and a scoped tool set, returning only its conclusion	E3
Tool surface	—	—	The set of tools/commands an agent can attempt to call; a smaller, well-named surface is easier for a model to reason about correctly	E3
Verification debt	—	—	Unverified agent output piling up between "it ran" and "it's right"; one of a self-running loop's four silent costs	E3
Verification ladder	—	—	The escalating rigor of a check — assertion, executed test, fresh-context review, cross-model audit, human gate; referenced as a planned companion page not yet published	E4

How to read the grades

Same four grades as the layers, applied per term instead of per layer:

Grade	Means
E1	Measured — a number, from a named corpus, reproducible by you
E2	Reported — a named source, a primary link, a date; not independently reproduced
E3	Argued — a coherent mechanism, no measurement attached
E4	Asserted — confident, popular, unsourced

A high grade doesn't mean "true," and a low one doesn't mean "wrong" — prompt engineering is E3 in this table and it's still a real, working practice. The grade tells you what you're entitled to assert without checking further, and where you'd have to follow the link yourself before repeating it as fact. Never let an E3 or E4 term write a permanent rule into your config; they're fine for deciding what to try, not for a line that will sit unexamined in every future context window.

---

See also: the layers · context · harness · the gate · orchestration · the five moves.