Comparison Changelog

The May 13 release finalized the substrate for off-turn bundle work: @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; and 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 the 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 editor; feature gating is uniform across bundle/api/mcp/widget/cron via canonical enabled.* flags.

Why it matters: the same substrate now supports long-running bundle workloads such as memory analysis and reconciliation. The current memory posture is explicit and user-reviewed: durable memories are inspectable records, snapshots protect restore paths, and reconciliation runs as a tracked async job. The honest comparison is in Claude Dreams-style memory vs. KDCube ReAct memory.

Release notes →

KDCube's memory model is being clarified around separate surfaces: ReAct conversation memory, internal memory anchors, indexed notes, durable user memory, ANNOUNCE hotsets, and snapshot-backed reconciliation. Durable user memory is not an automatic transcript rewrite; it is user-visible state with scope, provenance, and restore paths.

Why it matters for KDCube: this avoids mixing agent-authored internal notes with durable user preferences. The next meaningful product work is not a hidden "dream" pass; it is a widget-centered workflow for search, pinning, snapshots, async analysis, proposal review, apply, and restore.

Comparison post →

The "Build yourself" column was rewritten to reflect what teams actually stitch in 2026: Temporal and Hatchet have largely replaced Celery Beat for durable agent scheduling; Pydantic AI v1.85+ sits alongside LangGraph as a serious lightweight agent loop with first-class durability and Logfire-native tracing; Langfuse (acquired by ClickHouse in January 2026) is now the de-facto OSS observability standard; managed sandbox options expanded to include Northflank and Daytona alongside E2B and Modal (Firecracker / microVM-backed isolation is non-trivial to self-host).

Why it matters: KDCube's value is the integration of these layers. The DIY path's individual primitives keep getting better; what stays painful is the cross-layer plumbing — tenant-aware authz, hot-reload of bundles, atomic budget enforcement, and channeled streaming over Socket.IO + SSE + REST.

Temporal → · Hatchet → · Pydantic AI → · Langfuse →

AgentCore (GA late 2025; April–May 2026 wave added São Paulo, Frankfurt, Tokyo regions) now bills across five separate consumption meters: Runtime (vCPU + GB-hr/sec), Gateway (per-MCP-op + per-indexed-tool), Memory (per event + per record retrieved), Identity (per token issued), and Policy ($0.000025/auth request) — plus S3 for browser artifacts. Production-grade Browser runtime, Code Interpreter, and Memory (short + long-term) reached GA in the same wave.

Why it matters for KDCube: AgentCore continues to mature on the runtime axis, but the five-meter cost surface is hard to forecast and tightly coupled to AWS-only operational primitives (Gateway, Memory, Identity, Policy, CloudWatch). The "self-host on infrastructure you own with one budget surface" pitch holds; the "managed services lock you in" framing on Why is the right read.

AgentCore pricing → · AgentCore product page →

CrewAI v1.14.x (April 2026) shipped checkpoint TUI with fork/lineage, reasoning-token + cache-token accounting, and A2A (agent-to-agent) protocol docs. CrewAI AMP — the proprietary control plane on top of the MIT engine — now offers an on-prem option, with real-time dashboards, per-agent cost tracking, RBAC, SOC2, SSO, secrets, and PII detection.

Why it matters for KDCube: CrewAI is the closest competitor on the "open core + on-prem control plane" axis. Differentiation: KDCube's runtime is fully MIT (no Enterprise tier gate), bundles are surface-agnostic (chat / REST / iframe / MCP / @cron in one app), and pricing is your infrastructure cost — not "contact us." Watch this row.

CrewAI changelog → · AMP pricing →

Anthropic's Agent Skills standard (open-sourced December 2025) was adopted by 32 tools within months — VS Code, ChatGPT/Codex, Gemini CLI, Junie, Kiro, Goose — and Vercel's skills.sh marketplace now lists ~89k skills. It is a packaging standard, not a runtime: a Skill is a folder of prompts + scripts + metadata, not a hosting environment.

Why it matters for KDCube: KDCube bundles can wrap Skills — the two layers compose rather than compete. We may add a future row noting compatibility, but Skills doesn't change the Why-page positioning between libraries / managed services / self-hosted runtime.

The New Stack — Agent Skills →

Snapshot refreshed against the May-2026 KDCube state: split-container exec strategy, hardened ISO runtime, OSS frontend runtime config served from ingress, Tier 1 bundle docs, file-backed bundle authority. KDCube's scoring on Governance & Security and Economics columns unchanged; sandbox isolation row gains a stronger primary source.

Why it matters: the next external snapshot of KDCube vs alternatives won't surprise readers — the differentiating rows are still the same ones (per-tenant budgets, atomic admission, tenant-scoped audit, isolated execution).

KDCube changelog → · Engineering write-up: split executor & trusted supervisor →

OpenAI's developer post formally defined AGENTS.md as a cross-tool standard, building on the AGENTS.md spec released August 2025 and now stewarded by the Linux Foundation's Agentic AI Foundation. Adopted by OpenAI Codex, Cursor, Factory, Claude Code, and Gemini CLI.

Why it matters for KDCube: not a comparison-table row — but it's the file every coding agent reads first. KDCube added AGENTS.md at the website repo root in May 2026 to be discoverable to the same tool ecosystem.

developers.openai.com — Custom instructions with AGENTS.md →

The official MCP Registry launched in preview as a metaregistry — an authoritative upstream that downstream UIs (Smithery, mcp.so, Glama, MCPfinder) can anchor to. Backed by Anthropic, GitHub, Microsoft, and PulseMCP.

Why it matters for KDCube: KDCube bundles can host MCP servers. As soon as a public sample bundle exposing an MCP server is ready, submit it to the registry. The comparison-table row "Hosts MCP servers" stays a KDCube ✅; the new lever is discoverability via the registry.

MCP Registry launch announcement →

The AGENTS.md open standard was published as a cross-tool convention for AI coding agents, with collaboration across OpenAI, Google, Cursor, Factory, and others. Now stewarded by the Linux Foundation's Agentic AI Foundation.

Why it matters for KDCube: not a comparison row, but it set the convention KDCube would adopt eight months later (May 2026).

agentsmd/agents.md repo →

LangGraph reached 0.x with stable graph-state checkpointing, thread IDs, and a UI inspector (LangGraph Studio). Persistence works out-of-the-box across SQLite/Postgres backends.

Why it matters for KDCube: the "agent state survives a restart" row is no longer a KDCube ✅ vs LangGraph ✗ — both deliver, with different shapes (KDCube: timeline-first conv.timeline.v1 in Postgres; LangGraph: graph checkpointer with per-super-step state). KDCube still differentiates on tenant-scoping and immutability — see the comparison table.

LangChain release notes — Jan 22, 2024 →

AgentScope (Alibaba) published with built-in OpenTelemetry tracing, sandbox tool execution, and configurable isolation. Distributed multi-agent focus.

Why it matters for KDCube: closes some "Partial" gaps in the orchestration tab (sandbox isolation, observability) but the gap on tenant-scoping and reservation economics is unchanged.

arXiv 2402.14034 →

The AutoGen project split into two tracks: Microsoft's AutoGen v0.4 (rewrite, async-first) and AG2 (community continuation of the v0.2 line). DockerCommandLineCodeExecutor for sandboxed code remains in both.

Why it matters for KDCube: tracking-only — KDCube users who run AutoGen inside a bundle should pin the variant; both work as a BYO agent loop.

microsoft/autogen →