Terminology
Terminology
A glossary for the current public Workloop and Gambit story.
Company and product names
Workloop
Workloop is the primary public brand, company, product, and destination. Public
CTAs should point people to Workloop and getworkloop.com.
Gambit
Gambit is the open-source synthetic scenario and evaluation layer for
trustworthy agent behavior.
Gambit is a secondary OSS story, not the main company or product brand.
Agent-building terms
Agent
An agent is a structured unit of AI work with explicit instructions, inputs,
tools, and outputs.
Prefer agent language by default.
Scenario
A scenario is a representative situation that exercises agent behavior. It may
simulate a user, task, workflow, tool pressure, policy edge, or failure mode
that should not regress.
Prefer scenario over scenario deck unless you are explaining exact CLI
syntax or file references.
Grader
A grader is an agent-backed or compute-backed check that scores scenario data, a
transcript, an artifact, or a run against expected quality or behavior.
Prefer grader over grader deck unless you are explaining exact CLI syntax or
file references.
Deck
Deck is an older repo term for a Gambit agent definition. You will still see
it in file names, code, and older posts.
Prefer agent or agent definition unless you are explaining file names, CLI
syntax, config keys, code/API names, trace event names, or legacy repo
compatibility.
Snippet
A snippet is a reusable prompt fragment or instruction block embedded into a
Gambit agent definition.
Prefer snippet, prompt snippet, or reusable instruction language over card
unless you are explaining exact .card.md filenames, gambit://cards/*
compatibility, or the defineCard helper.
Action
An action is a callable piece of work an agent can invoke during execution. It
usually handles a focused task such as retrieval, computation, or I/O.
Schema
A schema defines the shape of inputs or outputs so agent behavior stays typed,
inspectable, and easier to verify.
Execution and review terms
Runtime
The runtime is the execution layer that runs agent work.
Gambit is the OSS runtime foundation. Avoid implying that Workloop and Gambit
are the same layer.
Trace
A trace is the execution record of what happened during a run. It helps teams
inspect, debug, and verify agent behavior.
Verification
Verification is the process of checking whether an agent behaved the way you
intended. Depending on the workflow, that can include tests, review, traces,
comparisons, or other evidence.
Synthetic scenario generation
Synthetic scenario generation creates realistic users, tasks, workflows, edge
cases, or expected outcomes that an agent should handle.
Eval-data validation
Eval-data validation checks whether synthetic or hand-authored scenarios are
good enough to trust as evaluation data.
Scenario quality
Scenario quality is the check that generated or hand-authored scenarios are
realistic, covered, novel, difficult enough, grounded in the source material,
not duplicative, and clear about the expected outcome.
Behavior evidence
Behavior evidence is the transcript, state, trace, artifact, permission, and
grader record that explains what happened during an agent run and makes failures
reproducible.
Workloop direction terms
Project intent
A project intent is the explicit assignment that defines why work exists, what
success means, what constraints matter, and when to escalate.
This term is central to the broader Workloop product direction. Public docs
should use it only when the surrounding workflow is available or clearly framed
as product direction.
AAR
AAR stands for after action report. It is the structured record of what
happened during execution and what should be learned from it.
For external-facing copy, execution record or work report may be easier to
understand than the acronym.
Legacy terms
These terms may still appear in older posts, code, or compatibility notes:
deckas the primary public categoryagent harnessagent harness frameworkverification layerwhen it appears without scenarios, graders, traces, and
regression evidenceaugmentation pathquality control layersystem of record- eval-only or grader-first language used as the whole public Workloop story
They reflect earlier framing, not the preferred current public terminology.
Last updated: May 3, 2026