Glossary¶
This page defines terms used in the repository. Source paths are relative to
agentic-workflows-v2/ unless stated otherwise.
Workflow runtime¶
Adapter
An implementation of the runtime execution interface. The CLI supports:
langchain, which compiles a workflow to LangGraph and is the default for named YAML workflowsnative, which runs the repository's dependency-light DAG executor
See agentic_v2/adapters/.
Agent
The function or model-backed worker that executes one workflow step. A step
names its agent with the agent: field.
See agentic_v2/engine/agent_resolver.py.
Agent tier
A number from 0 through 5 used to select model capacity, output limits, and available tools. Tier 0 is intended for deterministic or lowest-capability work, but the current resolver can still pass a tier-0 agent through the model client. A tier number alone does not prove that no provider path is involved.
Checkpoint
Saved workflow state that a compatible execution engine can use to resume work. Checkpoint support is an optional engine capability, not a guarantee for every adapter.
See agentic_v2/core/protocols.py.
Condition
A when: expression on a workflow step. The step runs only when the expression
resolves to true. A skipped step has no usable output, so downstream mappings
should use coalesce() when they also accept another source.
DAG
A directed acyclic graph. Workflow steps are graph nodes and depends_on
values are edges. The runtime rejects missing dependencies and cycles.
See agentic_v2/engine/dag.py.
Execution context
The state for one workflow run. It holds workflow inputs, step results, services, and lifecycle events.
See agentic_v2/engine/context.py.
Expression
A value inside ${...}. Expressions can read inputs, step outputs, and context
values. They support a restricted set of comparisons, boolean operations,
arithmetic, collection literals, and coalesce(). Imports and arbitrary
function calls are rejected.
See agentic_v2/engine/expressions.py.
Fan-out / fan-in
A dependency shape where several independent steps run after the same parent, then a later step waits for all of them.
Loop
Bounded repetition of one step. loop_until: defines the stop condition and
loop_max: limits how many times the step can run.
Step
One unit of work in a workflow. A YAML step has a name and agent, and may declare dependencies, inputs, outputs, conditions, artifact contracts, tools, and a bounded loop.
Workflow
A YAML file that declares inputs, steps, dependencies, and outputs. Built-in
definitions live in agentic_v2/workflows/definitions/.
See Workflow authoring.
Contracts and artifacts¶
Artifact contract
Optional validation attached to a step input or output. The current
code_artifact contract accepts source code as a non-empty relative-path map
or as complete FILE/ENDFILE blocks. It rejects unsafe paths, empty content,
and placeholder-only output.
See agentic_v2/artifact_contracts.py.
Contract
A typed boundary model shared by runtime components. Most public contracts use
Pydantic models under agentic_v2/contracts/.
Protocol
A Python structural interface. A class satisfies a protocol by providing the required methods and properties; it does not need to inherit from the protocol.
See agentic_v2/core/protocols.py.
Run record
The JSON representation of a completed workflow run. Run records support the dashboard, comparison, and offline evaluation routes.
See agentic_v2/workflows/run_logger.py.
Models and tools¶
Fallback chain
An ordered list of model identifiers. When a model is unavailable or a call fails in a retryable way, the router can try the next configured model.
Model provider
A service or local runtime that accepts model requests. The repository has backends for cloud providers and local runtimes. Availability depends on installed extras, credentials, and reachable endpoints.
See Configuration.
Model tier
See Agent tier.
Tool
A named operation an agent can call, such as reading a file or making an HTTP
request. The runtime filters tools by tier and by a step's tools: allowlist.
An empty allowlist disables tools for that step.
See agentic_v2/tools/ and agentic_v2/engine/tool_execution.py.
Tool approval
A policy decision applied before a tool with side effects runs. Approval behavior depends on the configured policy and provider.
Retrieval¶
ARP ships no retrieval (RAG) package or command. The agentic_v2.rag package
and its terms (chunking, embeddings, BM25, hybrid retrieval, vector stores) were
removed by ADR-057; the standalone
groundkit repository is the
successor.
Evaluation¶
Criterion
One named part of a rubric, with a weight and scoring guidance.
Dataset
Input cases and expected values used by an evaluation runner.
Evaluator
Code that scores a result. The evaluation package includes structural, pattern-based, standard, quality, and model-backed evaluators.
LLM judge
A model-backed evaluator that scores output against explicit criteria. Its score is evidence from that judge configuration, not a human verdict.
Rubric
A set of weighted criteria used to score output. Criterion weights must sum to 1.0 within the loader's tolerance.
Runner
Code that applies an evaluator to one or more dataset cases and collects the results.
See Evaluation framework.
Operations¶
No-LLM mode
Development mode enabled with AGENTIC_NO_LLM=1. Supported model-client paths
return deterministic placeholder text, which is useful for interface tests. It
does not make every subsystem deterministic.
OpenTelemetry (OTEL)
The tracing standard used by optional runtime instrumentation. Tracing is disabled unless configured.
Readiness check
An endpoint that reports whether the server is ready to accept work. In this
repository it is GET /api/health/ready.
Smoke test
A small check that confirms a basic path works. test_deterministic is the
recommended provider-free workflow smoke test.