Skip to content

Introduction

Zakardo is an agent-native full-stack application framework for building applications where AI agent behavior is expressed as structured, versioned, validated programs called plans, composed of governed, typed units of execution called capabilities, triggered by user intents.

Current AI agent frameworks treat agent behavior as an emergent property of prompt engineering and tool calling. This model has fundamental issues:

ProblemDescription
No inspectable programThere is no versioned, testable artifact describing what an agent will do before it does it
No governanceSide effects are fire-and-forget with no policy layer
No learning loopSuccessful behavior is not captured as a reusable artifact
No auditabilityLogs are unstructured and not replayable
No composabilityEvery framework defines its own incompatible tool interface
No frontend modelNo architecture exists for intent-driven, plan-aware UIs

Current frameworks ask: How do I give an LLM the ability to call functions?

Zakardo asks: How do I build, govern, validate, and trust the programs that agents execute?

Every Zakardo application runs through the same core loop:

Intent → Planner → Plan Draft → Enrichment → Full Plan IR → Validator → Execution Engine → Trace
  1. Intent — The user expresses a desired outcome in natural language
  2. Planner — An LLM-powered, policy-aware planner generates a minimal Plan Draft using backward chaining
  3. Enrichment — A deterministic 8-step pipeline converts the draft into a fully typed Plan IR with schemas, dependency graphs, and metadata
  4. Validator — 10 structural and policy checks verify the plan is safe to execute
  5. Execution Engine — A durable state machine runs the plan step-by-step with checkpointing, retry, and governed side effects
  6. Trace — An immutable, append-only record captures everything for auditing, replay, and learning
ConceptWhat It Is
IntentUser request (natural language, packaged action, or structured form) — the entry point
PlanStructured, versioned program of capability invocations with typed data flow
CapabilityTyped, versioned unit of execution with declared inputs, outputs, effect class, and boundary
Effect ClassWhat a capability does: READ, WRITE_INTERNAL, WRITE_EXTERNAL, DESTRUCTIVE
BoundaryWhere it executes: LOCAL, REMOTE, USER_ENV
TraceImmutable, replayable record of every execution event
Policy / GateGovernance rules that can pause execution for human approval
A2UIGoogle’s Agent-to-UI protocol — declarative, streaming frontend rendering
Plan LibraryStores and promotes validated plans through a trust lifecycle
FrameworkFocusZakardo’s Difference
LangChain / CrewAILLM orchestration and tool callingZakardo operates at the execution runtime layer beneath LLM reasoning. Plans are inspectable programs, not emergent behavior
Temporal / PrefectDeterministic workflow orchestrationZakardo plans can be generated dynamically by an LLM planner. Capabilities have governance, effect classes, and trust lifecycles
OPA / CedarGeneral-purpose policy enforcementZakardo’s policy layer is purpose-built for agent execution, using effect class and boundary declarations
LangSmith / ArizeLLM call observabilityZakardo traces are structural, replayable execution records that drive regression testing and plan promotion
PackagePurpose
@zakardo/coreCore runtime — everything from types to the Hono server
@zakardo/clientZero-dependency typed TypeScript client SDK (HTTP + SSE)
@zakardo/studioVisual dev tool — browser, playground, inspector
@zakardo/testingTest utilities — factories, mocks, matchers, test runtime