PageRegister
terms18
Noteattribution while it is checkable
Revised2026-07-24
The terms of art
The field is young enough that most of this vocabulary has a datable first use and a named author. That is worth recording while it is still checkable.
Every entry names who put the term into circulation and when. Where a term is older than its current usage, the attribution is for the sense used on this site, not for the word.
-
Loop Primary
One agent doing a single job on repeat: gather context, take an action, check the result, go again until a condition is met. The smallest unit of agentic work.
-
Harness Primary
The walls around a loop: the sandbox it runs in, the tools it can reach, the memory that survives between runs, and the gates that decide what "done" means. Agent = Model + Harness.
-
Software factory Primary
Many harnessed loops running at once, fed by a queue of work and drained through a review gate into production. Not a bigger agent — an org chart made of loops.
-
Dark factory Primary
A factory that ships code no human has read, verified only by other machines. The name is borrowed from lights-out manufacturing, where the lights are off because only machines are on the floor.
-
Lit factory Primary
The same pipeline with the lights left on where judgment lives: agents build, a human reads what comes out, and judgment moves upstream to product, design and architecture.
-
Comprehension debt Primary
The widening gap between how much code exists and how much any human still understands. It accrues silently, with the tests green, and comes due late.
-
Verification debt Primary
The accumulating cost of AI-written code that has not been meaningfully checked. The same creature as comprehension debt, seen from the accounting side.
-
Back pressure Primary
The rule that you can hand a loop only as much autonomy as you can cheaply and reliably verify, and not one inch more. Generation is a wide mouth; verification is a narrow neck.
-
Harness engineering Primary
The discipline of building and iterating the environment around the model rather than the code it produces — guides that steer before it acts, sensors that observe after.
-
Guides and sensors Primary
Feedforward controls that shape the agent before it acts (context files, skills, codemods) versus feedback controls that observe afterwards (linters, tests, review agents).
-
Computational vs inferential Primary
Whether a check is deterministic, fast and cheap (a type checker) or a probabilistic judgment call (an LLM reviewer). The distinction decides which checks may gate and which may only advise.
-
Ralph loop Primary
A coding agent in an infinite shell loop, reading the same prompt file each iteration, using the filesystem and git history as memory, starting each cycle with a fresh context window.
-
Oracle Primary
The thing that answers whether a loop's output is acceptable. An oracle earns a loop the dark only if it is cheap, runs at high frequency, and cannot be faked out by what it is checking.
-
Scenario-as-holdout Primary
End-to-end user stories kept outside the codebase, like a machine-learning holdout set, so the agents building the system cannot see or overfit the tests that judge it.
-
Satisfaction Primary
A probabilistic success metric — what fraction of observed runs likely satisfied the user — used in place of a boolean "the tests passed".
-
Inner and outer loop Primary
The inner loop is investigate, implement, test, report — the agent's work. The outer loop is deciding whether it was the right change at all, and carrying the consequences of being wrong.
-
Context rot Primary
The degradation of an agent's reliability as its context window fills with its own history. The Ralph loop's fresh context per iteration is a direct answer to it.
-
Cognitive debt Primary
A team's declining ability to reason about its own system, distinct from the code's quality. Named alongside "semantic diffusion" and complacency with AI code as the standing risks.