Agora is still at the concept stage; agent.json is in draft; ANP is nearing finalization; MCP has become a de facto standard.
These four technologies were at these respective stages when Shanghai Jiao Tong University integrated them into its taxonomy of agent protocols. It was in May 2025.
The taxonomy distinguished context-oriented protocols from those focused on inter-agent communication. It introduced a second level of segmentation, separating general-purpose protocols from specialized ones (the latter dividing, in the communication portion, into human-agent, robot-agent, and system-agent).
No progress for agents.json
Since then, agents.json has not seen a new version – the last one dates from February 2025. The project appears abandoned (nonfunctional demos, 404 documentation, expired Discord invitation, a YouTube channel that hasn’t been updated…). Wildcard, the American startup that initiated the project, still exists. It has pivoted to GEO (Generative Engine Optimization).
The protocol extends the OpenAPI specification to allow the definition of contracts guiding LLMs in API usage. These contracts contain one or more calls describing a result. A way to preserve non-determinism in task execution while constraining the use of tools.
The approach is stateless. The agents.json files, preferably hosted in a /.well-known folder, are exposed to LLMs as tools via a dedicated SDK.
A2A, entrusted to the Linux Foundation
Google announced A2A (Agent-to-Agent) in April 2025. A few weeks after the taxonomy was published, it entrusted the protocol to the Linux Foundation.
A2A enables communication between agents built on different frameworks. They can mutually discover their capabilities (via capability cards), negotiate their interaction modalities, and operate without exposing their internal state, memory, or tools. The communication is JSON-RPC over HTTP(S).
A W3C Working Group around ANP
ANP (AgentNetworkProtocol) had moved to v1 shortly after the taxonomy was published. Since then, the community behind it has led a W3C AI Agent Protocol Working Group, with contributors including Google, Huawei, and Microsoft.
A draft specification was published at the end of January. It outlines the three main building blocks of ANP: identity (based on the DID standard), as well as the description and discovery of agents. The dynamic negotiation of inter-agent communication protocols is based on natural language. Version 1 introduced a proposed P2P transactional framework and a human in the loop option.
AITP remains in draft
Since the taxonomy’s publication, AITP (Agent Interaction and Transaction Protocol) has stayed in draft. This Web3-oriented protocol originated from the NEAR Foundation, the creator of a Layer-1 blockchain. It is intended to allow agents to exchange all kinds of structured data (UI elements, forms, payment requests…). The latest updates indicate connections with the NEAR wallet. EVM and SOL wallets are on the roadmap.
ACP, now a building block of AGNTCY…
LangChain is the initiator of ACP (Agent Connect Protocol). The spec covers discovery, group communication, identity, and observability. It is now part of the AGNTCY initiative, which Cisco leads to create “a stack for the Internet of Agents” – and which has been under the Linux Foundation’s umbrella since July 2025.
… as AComp, merged with A2A
AGNTCY also leverages AComp (Agent Communication Protocol). It is also under the Linux Foundation’s umbrella, where it has merged with A2A. It is supported, among others, by AWS, Microsoft, Salesforce, SAP, and Snowflake. It owes its origin to IBM, which produced the reference implementation within the BeeAI framework.
Compared with ACP, rather than imposing strict specifications upfront, AComp focuses on the functional aspect. It is said to be simple enough not to require a SDK (standard HTTP tools are sufficient).
LMOS still aiming for W3C standardization
LMOS (Language Model Operating System) comes from the Eclipse Foundation. It implements the W3C’s Web of Things (WoT) architecture across identity, transport, and application layers, centered on the JSON-LD format.
The project features a Kubernetes operator and a router, integrated into a single runtime. It also includes a Kotlin-based language for building agents. It has not yet entered the W3C standardization process.
Agent Protocol has changed hands
The latest version (v1) of Agent Protocol dates back to 2024. That year, the foundation that created the protocol handed it to a startup developing a smartphone AI assistant.
Built on OpenAPI, Agent Protocol defines a unified interface for lifecycle management. It introduces abstractions such as runs (task executions), threads (multi-turn interactions), and stores (long-term memory).
Academically originated protocols that remained concepts
Shanghai Jiao Tong University had included, in its taxonomy, several protocols from the academic world that were then at the concept stage. None seems to have a major reference implementation today.
Among them, Agora, born at Oxford University. Its latest version dates to January 2025. It enables agents to create ad-hoc protocols based on YAML documentation.
With PXP (Predict and eXplain Protocol), stemming from an Indian technological institute, the domain is human-agent communication. The protocol involves a whiteboard system and a planner that ensures turn-taking in the discussion.
In the same area, there is LOKA (Layered Orchestration for Knowledgeful Agents), from Carnegie Mellon. Drawing on standards such as DID and VC (Verifiable Credentials), it implements a decentralized consensus system based on shared ethical rules.
CrowdES is a robot-agent protocol born at the Korea University of Science and Technology for Engineering and Technology (Gwangju, South Korea). Designed to manage group behavior, it includes a “transmitter” and a “simulator.” The former uses diffusion models to assign individual attributes (agent types, movement speed, etc.) based on spatial information extracted from input images. The latter generates trajectories and interactions through a state-change mechanism based on Markov chains.
The University of Liverpool has advanced work on the family of protocols known as SPP (Spatial Population Protocols). They enable robots to agree on a coordinate system, even if that system is arbitrary and their starting positions may be as well. Each robot can store one or more coordinates and analyze the distance to other robots during interactions. The distance calculation can rely on a “leader” to anchor the coordinate system.
