Reference · 6 min

Glossary

Canonical terms for Loopmaster execution units, state, checks, budgets, stop rules, loop layers, and vendor terms from Claude Code and Codex.

Use this glossary when you write a loop plan or ask an agent to review one. Loopmaster vocabulary is normative for this site; vendor terms are descriptive mappings to the cited docs, not claims that Anthropic or OpenAI use Loopmaster's taxonomy.

Core terms

Term	Meaning	Often confused with
Prompt	One instruction or input to a model. It can describe a job, but it does not repeat work by itself.	Loop.
Scheduled prompt	The same prompt re-run on a timer. The timer repeats the prompt, but each run may be amnesiac, ungraded, and unaware of done.	Loop or automation.
Automation	Work that runs without a human's hands at the moment of execution. Scripts, zaps, workflows, scheduled prompts, and loops can all be automations.	Loop.
Trigger	The time, event, webhook, command, or human request that starts one pass.	Loop or schedule.
Schedule	A time-based trigger. It can wake a loop, but it is not the loop.	Loop.
Workflow	A predefined path through steps. A workflow may contain loops, but it does not have to let an agent choose the next action.	Agent loop.
Agent workflow	A workflow where some steps are performed by a model. The model fills in steps inside a predefined path; it does not own the outer continue-or-stop structure.	Agent loop.
Loop	Repeated passes with carried state, explicit checks, a stop or escalation rule, and a harness or runner that owns repetition.	Prompt or trigger.
Agent loop	The Loopmaster loop pattern with agency inside the pass: an agent observes state, chooses actions, uses tools, checks results, and stops or escalates.	Workflow.
Autonomous system	One or more loops granted standing permission to act within bounds without per-action approval. Autonomy is a permission property, not a control structure.	Loop or unattended automation.
Harness/runtime	The system that owns repetition: model calls, tool execution, state, logs, limits, approvals, and termination.	Model or prompt.
State	The umbrella term for what a loop can use across time: transient context, durable working state, and cross-run memory.	Context window.
Check	The proof that work succeeded: tests, source URLs, file existence, screenshots, reviewer approval, or another explicit signal; checks succeed or go green, never "the pass passed."	Agent confidence.
Stop rule	The condition that ends or pauses the loop on purpose; Loopmaster names exactly four terminal stops: done, blocked, exhausted, and interrupted.	Error handling.
Approval gate	A required human pause before public, destructive, expensive, credential, production, or otherwise risky actions.	Notification.
Log/artifact	The durable record a human or later agent can inspect: transcript, report, pull request, screenshot, saved state, or summary.	Final answer.
Runner	A concrete implementation of the harness, such as a native agent mode, scheduled task, cron job, queue worker, workflow, or script.	Trigger.

Units of execution

Loopmaster's smallest-to-largest execution chain is: action inside turn, turn usually inside pass, and pass inside run. A pass usually contains one turn, but a harness may group several turns under one pass-level check.

Term	Meaning	Example
Action	One tool call or externally visible operation, such as a command, file edit, search, or API request.	Running `npm test` once is an action.
Turn	One top-level unit of agent work that starts from a user or harness directive and ends when the agent yields a result or question.	A Codex turn can start with "fix the failing auth test," run several tool calls, and end with an assistant message summarizing the change.
Pass	One traversal of a loop: load state, work, check, record, then choose the next verdict.	A docs freshness loop reads the source list, updates one stale entry, records evidence, and writes a pass-log entry.
Run	One lifecycle of a loop under one contract and one budget, from start to a terminal stop.	A three-pass link-audit run starts from a contract and ends done, blocked, exhausted, or interrupted.
Verdict	The recorded outcome at a pass boundary: continue, done, blocked, exhausted, or interrupted.	A pass log can record `continue` after fixing two links when three unchecked links remain.
Iteration	Non-normative and ambiguous: it means one repetition of whichever cycle the author is discussing, not a Loopmaster unit.	OpenAI's Codex agent-loop article uses iteration for model-inference-to-tool-call cycles inside a turn, while the Codex Goals guide uses iteration policy for deciding what to try next after each attempt.

State: three kinds, three lifespans

Term	Meaning	Example
Context	The transient working material inside a session: messages, tool output, loaded files, and model working context.	A test failure pasted into the current conversation is context.
Working state	Durable per-run files that let the next pass rebuild where the run stands.	A task contract, checklist, pass log, and git history are working state for a multi-pass run.
Memory	Cross-run knowledge deliberately kept to shape future work.	`AGENTS.md`, a decision record, or this glossary can be memory when future agents read it before acting.

Context answers "what am I doing right now," working state answers "where am I in this run," and memory answers "what should change every future run."

Contracts, budgets, and stops

Every run ends in exactly one terminal stop. continue is a pass verdict, not a terminal stop.

Term	Meaning	Example
Contract	The written operating agreement for a loop: target, check, budget, state, constraints, and blocked rule.	A `task-contract.yaml` can say "make `npm test -- auth` exit 0 without deleting tests."
Check	The proof that work succeeded, preferably automated and specific enough that the loop cannot talk itself into being done.	A release-note loop can check that `npm run build` exits 0 and the changed page contains the new version number.
Budget	The run's fuse, usually expressed in passes, time, tokens, spend, or a combination.	A bug-fix loop can stop after three passes or thirty minutes, whichever comes first.
Pass log	The durable per-pass record of what state was read, what happened, what check ran, what evidence exists, and what verdict was recorded.	A pass-log entry can link the CI run, list files changed, and record `done`.
Done	The terminal stop where the check succeeds with evidence.	A quick-start run is done when the promised command exits 0 and the evidence is recorded.
Blocked	The terminal stop where the loop needs input and stops with a specific question and what it learned.	A release loop is blocked when signing credentials are missing and it asks for the exact credential or approval needed.
Exhausted	The terminal stop where the budget is spent before the check succeeds.	A flaky-test loop is exhausted after three attempts if the failure remains and the contract allowed only three passes.
Interrupted	The terminal stop where a human or outer system stops the run before another terminal stop fires.	A user pressing Ctrl+C during a non-interactive goal run interrupts it.

Blocked is a success mode, not an error: stop-and-ask beats guess-and-drift.

Layers (Loopmaster taxonomy)

Task loop, goal loop, and steering loop are Loopmaster's own taxonomy, not Anthropic or OpenAI vocabulary. A layer name applies only to work that already qualifies as a loop under the five-part test; defining a layer is not the same as claiming Loopmaster operates that layer.

Term	Meaning	Example
Task loop	A loop that varies actions until one task's check succeeds or a terminal stop fires.	A coding agent edits files, runs tests, reads failures, and edits again until the auth test is green or the run stops blocked or exhausted.
Goal loop	A loop that varies tasks until the goal's stopping condition is met, blocked, exhausted, or interrupted.	A milestone loop chooses one task loop, waits for its verified result, then chooses the next task until the feature is complete.
Steering loop	A loop that varies goals by watching world state and deciding what goal loops should pursue next.	A weekly product-maintenance loop could review usage, breakage, and landscape changes to re-aim the next goal contracts; Loopmaster does not claim to run one without a checked steering-loop pass log.

Vendor term map

Checked 2026-07-10. This map is descriptive, not an endorsement claim: when a vendor doc says the cited thing, Loopmaster maps it to the nearest Loopmaster term. "Closest analogue" means the vendor does not define the Loopmaster unit directly.

Loopmaster	Claude Code (Anthropic)	Codex (OpenAI)	Sources
Action	Tool use or tool call; Anthropic's agent framing has agents dynamically directing their own tool usage.	A tool call requested by the model and executed by the Codex harness before its output is appended for the next model query.	[S5], [S2]
Turn	A `/goal` turn is followed by a small fast model check; if the condition is unmet, Claude starts another turn.	One conversation turn runs from user input to assistant message, and may include many model-inference to tool-call iterations.	[S1], [S2]
Pass	Closest analogue: one `/goal` after-turn completion-check cycle, one timed `/loop` firing, or one Stop-hook evaluation.	Closest analogue: one attempt under an active Goal, with the iteration policy deciding what to try next after each attempt.	[S1], [S3]
Run	A `/goal` lifecycle from activation until the condition is met, a turn/time bound is reached, `/goal clear` is used, or the process is interrupted.	A Goal's thread-scoped lifecycle as a completion contract that may end in success, pause or clear, interruption, budget limit, or a blocker requiring input.	[S1], [S3]
Check	The goal's completion condition, checked after each turn by an evaluator that judges what Claude surfaced and does not call tools.	The Goal's verification surface; completion must be evidence-based against files, tests, logs, benchmark output, generated artifacts, or other concrete evidence.	[S1], [S3]
Budget	A `/goal` condition can include a turn or time clause to bound how long it runs.	A Goal carries budget accounting; reaching the budget limit is not completion.	[S1], [S3]
Blocked stop	Anthropic's autonomy research observes Claude Code pausing to ask for clarification as an agent-initiated form of oversight; the `/goal` doc does not define a separate blocked state.	Codex Goals can stop on a blocker requiring user input, and strong Goals should define a blocked stop condition.	[S7], [S3]
Working state	Long-running Claude agent harnesses use progress files and git history to bridge sessions that start with no memory.	Closest analogue: Codex manages a growing thread context and context window, and Goals persist as thread-scoped state rather than global memory.	[S6], [S2], [S3]
Memory / instructions	The cited Claude sources define no Claude Code memory-file equivalent; Anthropic's long-running-agent guidance keeps cross-session knowledge in progress files and git history.	Codex reads and layers `AGENTS.md` instruction files before work.	[S6], [S4]

Task loop, goal loop, and steering loop are intentionally absent from this vendor map: they are Loopmaster taxonomy, not Anthropic or OpenAI vocabulary.