Juan Reyero's aweb gives AI agents mailboxes that outlast sessions

Why it matters

Aweb is betting that agent teams need durable identities and communication outside any single model or runtime. Its Claude Code setup also shows how much a shared protocol still depends on runtime-specific permissions and integrations.

aweb is a communication layer for AI agents built around a problem familiar to anyone running agents in separate sessions: a message can be sent, but the intended recipient may be offline or never see it. Founded by Juan Reyero, the project gives agents stable identities, durable mail and chat, and wake-up events intended to connect different runtimes and machines.

Reyero brings a career in engineering and company-building to the project. He spent nearly two decades at HP, working in Barcelona, San Diego and Palo Alto on image processing and software for large-format and 3D printers, then held engineering leadership roles at printer companies Xaar and Quantica. He also founded GreaterSkies, an automated personalized star-map service. Reyero's stated thesis is that an agent's identity, knowledge and working relationships should persist even when its model provider, runtime or machine changes.

Dr. Eugenie Regan is also identified as a co-founder. Her earlier career was in conservation science and research products: Springer Nature described Regan as its vice president of Research Solutions, leading products and services for researchers. That background puts a second kind of operational experience behind aweb's infrastructure pitch: building tools around how people and organizations exchange work.

Messages that outlast a session

Aweb's core design separates the message from the signal that tells an agent to look for it. The server stores mail and chat as durable state; an event alerts a connected runtime that something is waiting. The agent then retrieves the message by ID. If the recipient is offline, the message can remain available until it reconnects.

That distinction is central to the product. Session transcripts are tied to a particular runtime and can disappear from the working context when a session ends. Aweb instead aims to give agents a stable address and a communication history that survives changes in process or machine. Its aw command-line tool, HTTP API, MCP tools and event stream are designed to let different agent setups use the same underlying system.

A demonstration on aweb's site, recorded on September 5th, shows Reyero's proposed workflow: an agent running Claude Code sends a question to an agent running Pi on another machine. The Pi agent wakes, reads a file and replies; the site says the response came in nine seconds, and the complete demonstration runs 95 seconds. That is a demonstration of a specific two-agent exchange, not evidence of broader deployment.

The software is MIT-licensed and available on GitHub, and the service can be self-hosted. Aweb also separates its identity system, AWID, from message delivery. In the company's design, DNS-backed namespace authority, team membership certificates, server hosting and custody of agent signing keys can be controlled separately. Organizations can use aweb's hosted service or bring their own domain and team authority.

The wake-up still depends on the runtime

Aweb does not control every part of the handoff. Its documentation for receiving events says the runtime or orchestrator must provide a push channel, managed wake loop or polling routine. An event is a signal, not a guarantee that an agent has acted on the message.

The Claude Code integration illustrates the tradeoff. Aweb's current setup instructions require Claude Code to run with permission prompts disabled for channel messages to appear in the session. The documentation says that, in default auto mode, notifications can arrive without being surfaced to the agent. Pi has a maintained extension; other runtimes can use the event stream or poll for messages. These details put a practical boundary around the broad claim of runtime independence: the communication interface is designed to be shared, while reliable delivery into a working session still requires runtime-specific integration and operator choices.

Reyero's bet is that agent teams will need infrastructure that behaves more like durable organizational communication than a chain of prompts confined to one tool. Aweb is trying to make that layer portable and independently operable before teams settle on a single model provider or runtime. The near-term test is straightforward: whether developers find the identity and message handoff useful enough to install, configure and trust across the systems they already run.