Glossary

Agent

A runnable role with a model, instructions, capabilities, skills, and optional subagents, defined in a recipe’s agent YAML.

Recipe

The versioned Git package containing one or more agent definitions and the resources they share. Recipes are the open format for vertical agents, built on Pi. See Recipes.

Pi

The agent harness that gives a model its instructions, context, and tools and coordinates the turns of the agent loop. Pi can run a recipe locally without Introspection; the platform runs the same semantics in a managed runtime. See pi.dev .

Runtime

One immutable deployable version of an agent, with the identity, policy, bindings, and resources needed to run it. Customer runtimes pin a recipe version; managed platform runtimes such as Operator use code from an Introspection release. See Runtimes.

Operator

Introspection’s AI engineering agent for improving the agents running in a project. Operator investigates production evidence, coordinates expert review, tests focused Recipe changes, and can prepare a pull request in a registered repository. Each conversation is saved so you can review or continue the work later. See Operator.

Automation

Saved instructions and schedules for recurring work in a project. A scheduled automation starts a task for each occurrence, or posts into an existing task, and keeps links to its results in run history. See Automations.

Recipe proposal

A selected recipe change recorded with its target runtime, source commit, and rationale. Proposals can be pending, accepted, or dismissed. Recording or accepting a proposal does not merge or deploy the change. See Operator.

Runtime group

The stable identity and lineage connecting every runtime version of one logical agent. Environment lanes, patterns, and experiments operate across the group.

Environment lane

development, staging, or production. Lanes keep traffic and configuration separate. Staging and production each select one deployed version per runtime group; development routes to an attached local introspection dev session.

Binding

A scoped endpoint, variable, or write-only credential resolved into a task’s sandbox at execution time. Bindings keep environment-specific configuration outside the portable recipe. See Bindings.

Task

One durable unit of agent work in one isolated sandbox. An interactive task can span several runs while retaining the same live sandbox. See Tasks and runs.

Run

One turn of work within a task: a prompt or steering message, its streamed activity, and its settled result.

Runner

The SDK handle an application holds while operating a runtime. It creates and resumes tasks, streams run events, and exposes task files, shares, feedback, and resolved context. See the SDK overview.

Sandbox

The isolated, disposable execution environment created for exactly one task. Its working filesystem is temporary; saved artifacts and evidence persist.

Conversation

The immutable record of what an agent did: messages, model calls, tool calls, errors, feedback, and outcome. See Conversations.

Observation

A platform-generated note about behavior seen in one conversation.

Pattern

A cluster of recurring observations across conversations. Patterns point to evidence worth investigating; they do not determine the fix. See Observations and patterns.

Continual learning

The loop in which production experience becomes structured signals, candidate changes are evaluated, and validated learning is distilled into a new agent recipe version. Observations, patterns, judges, evals, and experiments help surface or validate what should change; the durable learning is the reviewed recipe change that returns to production.

Harness distillation

One use of recipe optimization. A coding agent uses a stronger model’s eval result as the target while it improves the recipe for a fixed cheaper or faster model. Harness distillation changes the recipe, not the model’s weights. It starts from an eval suite, while the distillation described under continual learning starts from production evidence. See Optimize a recipe.

Judge and judgement

A judge is a recipe-owned quality standard evaluated against completed conversations. A judgement is one result from that evaluation. See Judges.

Experiment

A controlled production comparison that routes sampled subjects between runtime versions in the same group and evaluates a declared goal. See Experiments.

Pin

One of three related selections, disambiguated in context:

• A recipe pin records the exact repository path and Git commit used by a runtime version.
• A lane assignment selects the runtime version served by staging or production.
• Task resolution selects a version once when the task is created and keeps it for that task.

Bindings and judges can also be scoped to one exact runtime version. See Runtimes.

Control Plane and Data Plane

The Control Plane stores organizations, projects, runtime configuration, bindings, and keys. A Data Plane executes agents and stores what their work produces: tasks, conversations, events, files, and metrics. See Platform.