Skip to content

Glossary

This glossary defines the core terms used throughout the Zakardo framework. These definitions are authoritative — all specifications, documentation, and code use these terms as defined here.

Intent A user or agent request that expresses a desired outcome. It may be natural language, a structured payload, or a hybrid of both. Intent is input — it is never executed directly. It is interpreted by the Planner.

Plan A structured, versioned, executable program composed of an ordered set of capability invocations, along with data bindings, gating checkpoints, and policy expectations. Plans are the core artifact of Zakardo.

Capability A typed, versioned, atomic unit of execution. The smallest unit of work in the Zakardo runtime. Capabilities declare their input schema, output schema, effect class, boundary, and interaction requirements.

Plan Draft The minimal plan format produced by the Planner (LLM output). Contains a goal, sequential steps with capability references, and simplified bindings. Not executable — must be enriched first.

Full Plan IR (Intermediate Representation) The enriched, fully typed, execution-ready plan format. Contains resolved schemas, typed bindings with path expressions, dependency graphs, metadata, and error policies. Produced by the Enrichment Pipeline from a Plan Draft.

Plan Enrichment The deterministic 8-step pipeline that converts a Plan Draft into a Full Plan IR. No LLM involved — pure computation.

Execution Engine The durable state machine that runs validated plans step-by-step. Checkpoints after every transition, handles retry, gates, interaction points, and outbox management.

Execution Context The runtime accumulator that holds all data flowing through a plan execution — plan inputs, step outputs, interaction responses, and system metadata. Referenced via path expressions.

Execution Mode The safety level for plan execution: SIMULATED (nothing runs), DRY_RUN (reads only), LIVE (full execution), TRAINING (execution with detailed traces).

Trace An immutable, append-only record of everything that happens during a plan execution. Contains structured entries for every event — step starts, completions, failures, gates, interactions, and outbox operations.

Effect Class What side effects a capability produces: READ (no mutations), WRITE_INTERNAL (app-internal mutations), WRITE_EXTERNAL (external side effects), DESTRUCTIVE (irreversible operations). Ordered by severity.

Boundary Where a capability executes: LOCAL (within the server process), REMOTE (external service), USER_ENV (user’s browser/device).

Interaction Spec A declaration on a capability that it needs user input mid-execution. Types: SELECT, COMPARE, REVIEW, INPUT, CONFIRM.

Policy A declarative governance rule that controls agent behavior. Policies define conditions, scopes, and actions (GATE, DENY, WARN, LOG). Evaluated at three points: pre-generation, pre-execution, and per-step.

Gate A policy-driven pause in execution that requires human approval to continue. The runtime sends a gate prompt to the frontend; the user approves or rejects.

Outbox A staging area for external side effects during plan execution. Effects are staged (not sent) during execution and only committed (actually sent) after all steps succeed. Prevents orphaned side effects from partial execution.

Planner The LLM-powered component that resolves intents into plan drafts, clarification questions, or conversational responses. Uses backward chaining and is policy-aware.

Backward Chaining The plan generation strategy: start from the desired output (goal), find capabilities that produce it, check if their inputs are available, and recursively fill gaps until all inputs are satisfied.

Plan Library The store for validated plans with lifecycle management. Plans move through states: CANDIDATE → SHADOW → PROMOTED → DEPRECATED (or QUARANTINED). Trust scoring drives promotion.

Trust Score A weighted composite score that measures a plan’s reliability. Computed from execution success rate, training pass rate, user approval rate, and other signals.

CANDIDATE — Newly generated, unproven plan. SHADOW — Being tested against training datasets in parallel with promoted plans. PROMOTED — Proven reliable; used as the first source for matching intents. DEPRECATED — Replaced by a better plan; no longer used for new intents. QUARANTINED — Failed reliability thresholds; removed from active use.

A2UI (Agent-to-User Interface) Google’s protocol for agent-driven UI. Capabilities emit declarative A2UI messages that stream to the frontend and render using a component catalog.

Surface A named region in the frontend where A2UI content renders. Standard surfaces: assistant, workspace, notification.

Anchored View A pre-registered, persistent A2UI layout tied to specific capabilities or plans. Provides a polished, consistent experience for common interactions.

Ephemeral View A dynamically composed A2UI layout assembled at runtime for novel plans that have no pre-designed layout.

When discussing Zakardo, use these terms — not the alternatives:

Do NOT useUse instead
ToolCapability
WorkflowPlan
AgentBe specific: Planner, Execution Engine, or Runtime
StepCapability Invocation or Plan Step
Prompt (for input)Intent
WidgetA2UI Component
Page / Screen / RouteAnchored View or Ephemeral View