Competitive analysis

Contextual input delivery

Reviewed on 2026-09-08. The gaps below describe this package's delivery requirements, not verified defects in the compared libraries.

Reference	Reusable pattern	Requirement for this kernel	Decision
LangGraph interrupts	Keep resumable input state separate from its host-facing payload	Natural wording must not change the pending input	Add an optional presentation field, not a new routing path
AI SDK loop control	Bound generation and configure each step explicitly	Framing must not add an unbounded generation loop	Reuse existing composition and one repair
Mastra processors	Validate generated output before delivery	Required copy and protected content retain precedence	Reuse grounding review, exact-question validation and canonical fallback

Acceptance: existing interactions retain their default behavior; contextual wording preserves the required goal once, leaves input authority unchanged, honors redaction, and falls back after unsupported drafts. The packaged TypeScript contract is checked in an isolated consumer.

Runtime architecture

Reviewed on 2026-09-03 against the maintainers' documentation.

Library	What to reuse	Defects or gaps for this use case	How Intention Kernel differs
LangGraph	Explicit graphs, resumable state and interrupts	It is intentionally a low-level graph runtime; it does not define portable intentions, fact lineage, business-policy evaluation or effect receipts	LangGraph remains a private scheduler while the public contract is provider-neutral and capability-first
Mastra	Composable agents, workflows, memory and observability	Its broad application framework surface couples consumers to Mastra agents, tools, storage and runtime concepts	The package exposes one small kernel facade and injects models, durability, domain ports and events through interfaces
Semantic Kernel	Semantic operation descriptions, dependency injection and automatic function calling	Provider function calling can blur the boundary between model proposals and authorized side effects	Models propose intentions only; the planner validates capabilities and policies, and every write crosses confirmation plus a durable effect ledger
XState	Typed actor inspection, nested behavior and explicit transition events	Finite statecharts require enumerating paths and do not provide semantic intention interpretation or evidence grounding	Dynamic LLM proposals become a validated DAG while events retain step-level causal branches and nested workflow visibility

Resulting quality requirements:

• LangGraph and provider tool-call types never appear in a supported import path.
• The model cannot mutate state, execute a capability or authorize its own write.
• Multiple intentions are represented explicitly and planned by dependency, not flattened into a single route.
• Facts are versioned, evidence-bearing and invalidated through declared lineage.
• Model, durability, event and domain adapters are instance-scoped and swappable.
• Nested workflow engines remain adapter details and publish progress through one sanitized causal event contract.
• Package verification installs the tarball into a clean consumer before release.
• Expected ambiguity and missing information remain structured interactions, not exceptions or hidden deterministic guesses.