Glossary
Terms that are easy to mix up.
| Term | Means | Not to be confused with |
|---|---|---|
| Pipeline | A DAG of stages defined in *.pipeline.yaml |
An agent |
| Run / Pipeline run | One execution of a pipeline. Listed on the Pipeline runs page | An agent session or agent run |
| Stage | One step of a pipeline, executed by a model | An agent |
| Agent | A project-owned conversational definition in agents/<id>.agent.yaml that answers people over time. See Agents |
The stage agent: field (below) |
| Agent backend | The stage field agent: pi | claude and the agent model.backend that choose which model runtime executes |
An Agent |
| Agent session | A conversation with an agent (many turns). Listed under Agent sessions | A pipeline run |
| Turn | One message in and one reply out within a session | A stage |
| Agent run | The run inspector's view of one turn: steps, waterfall, messages, logs, configuration | A pipeline run |
| Revision | A saved version of an agent definition (v3). Sessions are pinned to the revision they started on |
A git commit |
| Approval | A paused ask-first tool call waiting for a person | A pipeline gate |
| Gate | A pipeline stage waiting for a human answer (HITL) | An agent approval |
| Inbox | The console's "Needs you" page: gates and agent approvals | |
| Connection | A named external service for channels (AgentMail, Slack, Telegram, iMessage, WhatsApp) in connections.yaml. Agents answer on one; stages send on one |
An email account, an MCP server |
| Channel | The use of a connection by a stage (channels:) or an agent (channel:) |
A connection |
| Memory | A host folder (memories/<id>/): memory.yaml version 1, a cover page (body.md), and cards beside it. A pipeline or an agent stores the memory id and, on Pi, can read and replace the cover and add, update, read, and search cards |
A card key, a run artifact, a session transcript |
| Memory card | A keyed note in a memory. The key is the subject's own id (an email, a ticket id, a version, or a path), 1–1024 characters after trim. A Pi stage or agent session can add, update, read, and search it. The operator can correct it or delete it | The memory id, the cover page |
| Email account | An IMAP/SMTP mailbox in email.yaml, used by email triggers |
An AgentMail connection |
| Web chat | The public /chat/<slug> page for an agent |
The Workshop chat |
| Workshop | The console's pipeline-authoring chat | Web chat |
| MCP server (Settings) | The Stageflow MCP endpoint other tools use to drive Stageflow (Settings → MCP server) | Project MCP |
| Project MCP | Tool servers (from .mcp.json or saved in the console) that agents and stages can use |
The Stageflow MCP endpoint |
| Check | A probe that starts a Project MCP server in isolation and lists its tools | A connection Check now (a live provider check) |
| Console secret | A write-only value saved in Settings → Project MCP that fills ${NAME} in MCP servers |
A stage secrets: entry from the Host environment |
| Grant | A tool an agent may call (tools:), enforced on every call |
An approval |
| Ask first | A grant that needs a person's approval each time the call is new in a session | A block |
| Eval | The file evals/<id>.eval.yaml that names one stage or one agent, a dataset, graders and run defaults. See Evals |
An eval run, or an agent's evals: groundedness setting (Agent observability) |
| Dataset | A file of cases (datasets/<name>.dataset.jsonl, .json or .csv) |
A task input |
| Case | One test input in a dataset: id, input, optional expected, tags. For an agent eval it is a scenario |
A result |
| Repeat | One execution of a case. Each case runs repeats times because models vary |
A retry of a failed stage |
| Result | One (case, repeat): the stage output and its grades | A pipeline run result |
| Grader | One rule that scores a result: schema, field, reference, rubric or code; for agents also tool_calls, conversation_check or goal_reached. Its verdict on one result is a grade |
A gate or a verify check |
| Rubric | A grader written in words and judged by a model | A rubric for human reviewers |
| Judge | The model that applies rubric graders (anthropic, openai or openrouter models) |
The stage's own model |
| Eval run | One execution of an eval: all results, grades, statistics, cost and a config snapshot (er_… ids, Run #n) |
A pipeline run. The stage runs it starts are marked origin = eval and hidden from the Runs list |
| Scenario | The input of a dataset case when an eval targets an agent: a list of user messages (script) or a persona, goal and opening (simulated), plus optional facts and tool mocks. See Evals for agents |
A pipeline task |
| Driver | What plays the user in an agent eval: a script of fixed messages, or the simulator | A pipeline driver or a model backend |
| Simulator | The model that plays the user in a simulated scenario (defaults.user_model, default the judge) |
The judge, which grades; or the agent under test |
| Conversation subject | What agent-eval graders read: the messages, recorded tool calls, final reply, turns, stop reason and usage of one conversation | A stage's envelope |
| Tool policy | How an agent eval treats tool calls: read_only (default; write, shell and unlisted tools are stubbed and recorded), mocked or real (needs explicit consent). See Evals for agents |
An agent's grants or its ask-first list |
| Eval session | The hidden agent session an agent eval opens for each conversation. See Evals for agents | A session a person started |
| Config snapshot | Everything that determined an eval run, stored with it and never edited | The eval file, which can change later |
| Fingerprint | The first 8 hex characters of a hash of the config snapshot. Equal fingerprints mean the same setup | A git commit |