Glossary

The words, before they are used on you. In reading order rather than alphabetical — each one stands on the ones before it, so ten minutes here reads as one story: a model, a loop around it, a program running the loop, and a record of what happened.

model

The neural network underneath everything else: text goes in, text comes out, and it remembers nothing between calls beyond what is sent to it again. Everything this site describes is machinery arranged around that one fact.

token

The unit a model reads and writes in — a short fragment of a word — and therefore the unit everything is billed in. Every cost figure on this site is, underneath, a count of these.

ConceptToken economics and context budgets

context window

How much a model can hold in front of itself at once, measured in tokens. Everything the agent knows on a given turn has to fit in it, is re-sent on every turn, and is paid for on every turn — which is why so much of the engineering here is about deciding what deserves the space.

ConceptContext engineering

agent

A model run in a loop with tools: it reads, acts, sees what its action did, and acts again until something stops it. The model does the reasoning; the loop is what makes it an agent.

ConceptAgentic engineering

harness

The program wrapped around the model that runs that loop — feeding it files and tool results, applying its edits, carrying its history. When a coding agent edits your repository, the harness is the thing actually touching the disk.

ConceptAgentic engineering

harness thinking

The working idea this whole guide is organised around: the engineering that decides whether agent work can be trusted lives in the harness and the rings outside it — what the model is handed, how its claims are checked, what gets written down — not inside the model. The model is the one part you did not build.

ConceptAgentic engineering

orchestrator

The layer above the agents: it decides what gets worked on next, starts and stops sessions, enforces the limits, runs the gates and keeps the record. Conductor, the tool whose record this site publishes, is one.

ConceptMulti-agent orchestration

session

One continuous stretch of an agent working — from being handed a prompt to finishing, hitting a ceiling, or being stopped. The session is this site's unit of account: costs, outcomes and durations are all recorded per session.

run

The whole undertaking: one plan carried out across however many sessions it takes, over hours or days, surviving restarts. A run is what leaves a record worth publishing.

ConceptDurable execution and resumability

plan

A declarative file stating the work before any of it starts: the stages, the checkpoints inside them, the gates that verify, the limits on money and tokens. The orchestrator enforces what is written rather than re-deciding it.

checkpoint

One deliberate break-point in the plan — a piece of work small enough to finish, verify and record before moving on. The tracker's table of checkpoints is the run's contract, and "done" is claimed by the agent but only confirmed by the engine.

gate

A command with an exit code, run by the orchestrator to check the work — a build, a test suite, a linter. It passes or it does not, and no one asks the agent's opinion. Everything else is an opinion with a schedule.

ConceptEvals, gates and acceptance

ceiling

The cap on what one session may spend, set in tokens. The nudge — the polite request to wrap up as the ceiling approaches — is its cooperative half, and whether agents actually honour it is one of this site's measured questions.

ArticleThe nudge that sat below the median

rollover

A session that hit its ceiling with the work unfinished, its remainder handed to the next session. Cheap when the record survives the boundary; expensive when it does not.

ArticleThe ledger that lied

handover

The note one session leaves for the next: what was done, what is left, what to watch for. Continuity between sessions is written, not remembered — nothing else survives the boundary.

ConceptAgent memory

evidence

An artifact recorded to prove a claim — a log, a screenshot, a figure recomputed from the store. On this site the rule is structural: prose names an evidence key and the build fetches the value, so a number can never be typed into a sentence.

ConceptAgent observability

owner

The human the machine stops for. An owner gate is a place the run pauses on purpose and says exactly what it is waiting for and what would clear it — which is what "human in the loop" means once it is a mechanism rather than a slogan.

ConceptHuman in the loop