Standards tracker
A standing record of the standards, runtimes, and protocols that constitute the agentic stack as the Review sees it. Each entry has a status, a version, a maintainer, an editorial note, and a link to a longer analysis. Updated as the working groups update.
| Standard | Status | Ver. | Updated | Steward | Editorial note |
|---|---|---|---|---|---|
| AutoGen | Active | 0.4 | Microsoft Research | A research framework that has cleaned itself up significantly for production. The conversation-based model is its own design choice, not the only choice. | |
| Composio | Active | 1.2 | Composio | A tool-catalog provider rather than a protocol, but the closest thing the category has to an "App Store for agent tools." Increasingly an MCP citizen. | |
| CrewAI | Active | 0.9 | CrewAI | The friendliest of the major frameworks for developers without a multi-agent background. Production maturity has improved sharply in the past six months. | |
| LangGraph | Active | 0.6 | LangChain | Not a standard, but a runtime that has become a de facto reference for how to model agent state. Worth tracking here because the shape is being copied. | |
| Letta / MemGPT | Active | 2.1 | Letta Labs | Less a protocol than a memory architecture, but everyone in the category is converging on the rough shape Letta sketched. | |
| MCP | Active | 1.0 | Anthropic, public working group | MCP is the closest thing the category has to a winning interop standard for the tool-and-data layer. Adoption has crossed the threshold where new clients are expected to support it. | |
| Phidata | Active | 2.5 | Phidata | Smaller surface area than the LangGraph / AutoGen / CrewAI cohort. Good for the workflow-and-data slice; less so for general orchestration. | |
| A2A | Draft | 0.4 | Google, public working group | The draft has improved meaningfully in the past two quarters. The political question — does a Google-led standard land in this layer? — is still open. | |
| AEB | Draft | 0.3 | Academic consortium | We are skeptical of benchmark suites in this category by editorial policy, but we cover the conversation because the standard's design choices are themselves interesting. | |
| Agent ID | Draft | 0.2 | Joint working group (informal) | The standards conversation that everybody knows is coming and nobody has yet been able to convene with sufficient authority. This is the placeholder draft. | |
| OpenAgents | Draft | 0.7 | OpenAgents Working Group (independent) | The most credible of the vendor-neutral attempts. Reads like it was written by people who'd built one of these before. | |
| TCT | Draft | 0.1 | Proposed by The Agentic Review | Our own proposal. We do not generally propose standards — we cover them — but this gap is large enough that we wrote a draft to put a position on the table. | |
| CapsM | Deprecated | 0.1 | Withdrawn — superseded by OpenAgents and A2A drafts | Documented here because the design conversation it started is still alive in OpenAgents and A2A. The artifact itself is deprecated. |
What this tracker is and isn't
The tracker is a reference, not a leaderboard. We rank by status and alphabet, not by our opinion. The opinion lives in the linked analysis pages and in the long-form articles cross-referenced from each entry. We do not score standards on a 0-to-100 scale or assign them quality ratings — we describe what they are, where they are in their lifecycle, and what we are watching about them.
The status taxonomy is deliberately coarse. Active means the standard is in use in production and the spec is stable enough to write against. Draft means a working group is iterating on it and consensus has not yet settled. Deprecated means the standard has been superseded, withdrawn, or is no longer maintained.
We update the tracker when (a) a spec releases a new version, (b) the steward changes, (c) the status changes, or (d) our editorial position on the standard shifts materially. Each update is dated on the entry.
If you maintain a standard that belongs on this tracker, write to the editors. We do not accept paid placements; we add entries that meet our coverage threshold (at least one production user we can name, an open specification, and a maintainer willing to answer questions).