Sources. Each cell's tooltip names a primary source — typically a release-notes page, a feature changelog, an authoritative repo path, or a documented spec. KDCube cells link to the platform repository; competitor cells link to the competitor's primary source for the claim. Where a feature exists but is gated (Enterprise tier, paid plan, beta), it's marked Partial with the gating reason in the tooltip.

Ratings.

• ✅ Yes — feature ships built into the system and works without writing glue code or paying for a separate tier.
• Partial — feature exists but with a meaningful gap: paid-only, beta, soft enforcement, or scope-limited.
• ✗ No — not a feature of the system. Often by design (e.g. agent-loop libraries leave hosting to you).
• DIY — Build-yourself column only. Names the most common open-source glue you would use, with rough engineering effort where known.
• Roadmap — KDCube column only. Planned, not shipped — typically for the managed KDCube.cloud service.

Cadence. The full table is re-audited on every major KDCube release and at least quarterly for competitors. Per-feature changes — including "competitor X added Y on date Z" — are recorded in the comparison changelog with their primary sources, so the diff between snapshots is auditable.

Scope of comparison. The four tabs map to four kinds of project we consistently see KDCube evaluated against:

• Orchestration libraries — LangGraph, CrewAI, AutoGen / AG2, AgentScope, Dify, Flowise.
• Cloud agent services — Bedrock AgentCore, Vertex AI Agent Builder, Azure AI Foundry / Agent Service.
• Dev-tooling — Claude Code, Cursor, GitHub Copilot, Devin (where they overlap with KDCube's bundle-authoring story).
• LLM app platforms (PaaS) — Dify, Flowise, Langflow, Botpress.

What changed in the May 13, 2026 snapshot. Aligned with KDCube release 2026.5.13.117: @cron now carries a span dimension (system | process | instance) gated by canonical enabled.cron.<alias> flags; a proc-owned scheduler with per-job Redis locks and a reconcile loop landed alongside selectable scheduler backends; a new bundle background job stream (kdcube_ai_app/infra/jobs/stream.py) gives off-turn jobs a generic async progress/log channel addressable from the admin UI and chat surface. Bundle props now reactively call an on_props_changed hook; dynamic per-resource config overrides (expr_config, tz_config) ship with an in-place admin override editor; and feature gating is now uniform across bundle/api/mcp/widget/cron via canonical enabled.* flags. On memory, the honest comparison is now explicit: Dreams-style systems expose off-turn memory consolidation as a product primitive; KDCube ReAct keeps conversation memory anchors, indexed notes, and durable user memories as separate surfaces, with durable memory maintenance snapshot-backed, inspectable, and user-reviewed. See Claude Dreams-style memory vs. KDCube ReAct memory.

What changed in the May 10, 2026 snapshot. The DIY column was refreshed to reflect the 2026 stack: Temporal / Hatchet have largely replaced Celery Beat for durable agent scheduling; Pydantic AI v1.85+ now sits alongside LangGraph as a serious lightweight agent loop; Langfuse (acquired by ClickHouse in January 2026) is the de-facto OSS observability standard; and managed sandbox options now include Northflank and Daytona alongside E2B and Modal. AWS Bedrock has been clarified as AgentCore (GA late 2025; five separate consumption meters — Runtime, Gateway, Memory, Identity, Policy). Vertex AI is the Agent Builder family (rebranded around Gemini Enterprise at Cloud Next 2026). New entrants worth noting but not yet given full table columns: Anthropic Agent Skills (an open packaging standard, not a runtime — KDCube bundles can wrap Skills) and the Microsoft Agent Framework (the AutoGen v0.4 + Semantic Kernel successor for MS-shop buyers).

Disagree? Push back on any cell on GitHub Discussions — include the source URL and we'll either fix the row or update the tooltip with your reading. The point of the comparison is to be wrong out loud, not silently wrong.