Glossary
Every term with its coiner, where one exists. Where a term has no single origin, this page says so explicitly rather than leaving the line blank — an absent attribution reads as an oversight, and provenance is part of the claim.
- Agent = Model + Harness
-
The decomposition that makes harness engineering a discipline. The model is bought or rented and changes underneath you; the harness (context, tools, memory, gates) is what you actually build and own.
- Back pressure
-
You can hand a loop 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.
- Comprehension debt
-
The widening gap between how much code exists and how much any human still understands. Unlike technical debt it accrues silently, with the tests green, and it is invisible until something breaks in a place nobody can read.
- Dark factory
-
A pipeline that ships code no human has read, verified only by other machines. The word is a claim about where reading happens — nowhere — not a claim about rigour.
- Deterministic floor
-
The binary, cheap, non-negotiable checks a change must pass before anything expensive looks at it: it builds, the types check, the links resolve. A floor is never a taste proxy, and the moment it starts approximating quality it stops being trustworthy.
- Digital Twin Universe
-
Behavioural clones of third-party services, built by agents from public API docs and shipped as self-contained binaries, so scenario tests can run thousands of times an hour without rate limits.
- Harness engineering
-
Building and iterating the environment an agent runs inside: guides that steer before it acts, sensors that observe after. The human's job moves from writing the code to improving the harness that writes it.
- Inferential sensor
-
A feedback control whose verdict is a judgement: an LLM-as-judge, a semantic review agent. Slow, non-deterministic, expensive, and the only kind that can see the things a linter cannot.
- Light factory (lit factory)
-
The same pipeline with the lights left on where judgment lives. Agents still do most of the building; a human reads what comes out and, more importantly, spends judgment upstream on product, design and architecture before a loop starts.
- Loop
-
One agent doing a single job on repeat: gather context, act, check, go again. The smallest unit of agentic work; everything above it is loops stacked on loops.
- Ralph loop
-
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. Named for the Simpsons character: the dumb persistence is the feature.
- Satisfaction
-
A probabilistic success metric (what fraction of observed trajectories likely satisfy the user) replacing the boolean of 'tests pass'. Accepting it means accepting that correctness is now a distribution.
- Scenario-as-holdout
-
End-to-end user stories kept outside the codebase, like a machine-learning holdout set, so the agents that write the code cannot see, overfit or game the tests that judge it.
- Spec-driven development (SDD)
-
Treating a written specification as the executable source of truth an agent works from, rather than a chat transcript. The on-ramp most teams take into factory work.
- Steering loop
-
The outermost of three nested loops: the model reasons and acts, the agent self-corrects against sensors, and the human iterates the harness. 'Keep quality left.'
- Verification debt
-
The accumulating cost of AI-generated code that has been merged but never properly checked. The sibling of comprehension debt: one is what nobody understands, the other is what nobody verified.