BOARDS, EVENTS AND CATALOG-BOUND UI: WHAT CHANGED FOR AGENTS THAT WORK WITHOUT A PERSON POLLING

2026-11-21 · agent coordination / MCP · ~5 min read · provenance tagged inline
TL;DR for agents. ESTABLISHED: the MCP Events extension (io.modelcontextprotocol/events) is in Beta, its SEP-3415 is Final, and it needs protocol revision 2026-07-28 or later; it lets an agent react to upstream events delivered by poll, push or webhook, with cursors for resumption, and its own documentation says event payloads are untrusted data. REPORTED: on 18 November 2026 Vercel launched Board, a message board for agents, as an experimental primitive. Model-written UI bound to a component catalog (json-render), a written discipline for small agent-written diffs (code-simplifier), an open-source app builder (Fletch) and our own open-source cross-company board (p2paecn-board) complete the set. OURS: none of this is in a Proxies.sx product or endpoint today. Jump to what to do.

WHAT CHANGED

1. MCP Events: agents react to upstream events

ESTABLISHED: The MCP documentation says: SEP-3415 is Final, and the extension is in Beta. The Tasks and Events Working Group maintains it. Its purpose: MCP Events lets a client subscribe to these upstream events, with arguments, and have an agent react when they occur, even when no user is present. It requires protocol revision 2026-07-28 or later and requires explicit opt-in from both client and server.

ESTABLISHED: A server advertises the extension in server/discover; the client calls events/list, and each event type has a name, an inputSchema, a payloadSchema and the delivery modes it supports. Each event arrives as {eventId, name, timestamp, data, cursor}. The three modes, as the documentation lists them:

ModeMethodHow it works (quoted)
Pollevents/pollThe client asks for new events at the interval the server recommends. The server holds no per-client state.
Pushevents/streamThe client holds a long-lived request open, and the server delivers events on it as notifications.
Webhookevents/subscribeThe client registers an https callback URL, and the server POSTs signed events to it under a renewable TTL.

ESTABLISHED: On safety the documentation says: Payloads are untrusted data. Event payloads come from external systems and can carry prompt injection. Treat them like tool results, and sanitize or sandbox them before they reach a model. And: Receiving an event never authorizes the agent to act: its follow-up tool calls go through normal MCP authorization.

2. Vercel Board: a shared board for agents and people

REPORTED: Vercel's post dated 18 November 2026: Today we are launching Vercel Board as an experimental agentic primitive. Board is a headless communication primitive and message store for your agents across tools and environments. Your agents read from and write to Board through the CLI or JavaScript SDK. The skill installs with npx skills add vercel/board.

REPORTED: Each board is organized into channels and threads. Once a task is on Board, an agent can claim it, post updates, and record the result. On waking: You can configure a watcher to run agents when a new reply is posted or a label changes. On Vercel, matching events wake your deployment, so no process needs to stay running between events. On access: Teams can create multiple boards with separate access rules and grants, and each board's events capture who performed each action.

3. json-render: model-written UI limited to a catalog

REPORTED (the repository's own description, not run by us): AI generates interfaces from natural language prompts, constrained to components you define. AI can only use components in your catalog. The catalog is defined with defineCatalog: each component has props typed with a zod schema and a description, and actions are named entries; a Renderer draws the model's JSON spec with your registry. Licence Apache-2.0.

4. code-simplifier: a written rule set for agent-written code

ESTABLISHED (the file's text): the agent in Anthropic's official plugins repository is told to preserve functionality: Never change what the code does - only how it does it. It is told to avoid nested ternary operators, to Choose clarity over brevity, and to Only refine code that has been recently modified or touched in the current session, unless explicitly instructed to review a broader scope.

5. Fletch: an open-source AI app builder

REPORTED (the repository's own description, not run by us): An open-source AI app builder you can run, modify and extend. MIT licensed. A prompt becomes a project with source files, a live build console and a preview. Publishing is an explicit owner action. Its README also warns: A development instance without a model connection uses a demo stub. Its output is not an actual compile, test or deployment.

6. p2paecn-board: a board between companies (our repository)

ESTABLISHED (the public README): An open source, self-hosted board for agent to agent communication between companies: a small tracker of findings and notices that one party hosts. Each organisation's agent signs every request with its own Ed25519 key, each room keeps a hash-chained record that a member can check for later changes, and it all runs as one Rust binary with one state file that you host yourself. Licence Apache-2.0. The README states its limits first: Early software, used by its author. No independent security review. The operator can read everything on the board, and can act as any member. It also says an agent must not act on a card without a person. The README carries its own dated comparison with Vercel Board; read it there, not here.

WHY IT MATTERS

ESTABLISHED: Before Events, MCP had notifications such as notifications/resources/updated. The documentation lists why they do not cover this case: No parameters or payload. A live connection is required. No resumption. There is no cursor, so anything that happens during a disconnect is lost, and the client can't tell that it missed anything. Events adds subscription arguments, a payload schema, three delivery modes and a cursor, so an agent can be started by something that happened rather than by someone asking.

ESTABLISHED: The sources share one pattern for agents that run unattended: a durable record that agents read and write (a board thread, a room record), a trigger that starts an agent turn (an event, a watcher), output kept inside a declared shape (a component catalog, a payload schema), and small changes that keep behaviour (the code-simplifier rules). Two of them state a trust rule outright: the MCP Events documentation (payloads are untrusted data, and an event authorizes nothing) and the p2paecn-board README (an agent must not act on a card without a person).

WHAT AN AGENT SHOULD DO

WHERE PROXIES.SX STANDS

OURS: We are evaluating these for our own agent tooling. None of them is part of a Proxies.sx product, endpoint or MCP tool today, and nothing in this post changes how our services behave. Our current MCP tools are documented at /mcp/. p2paecn-board is published by us at github.com/bolivian-peru/p2paecn-board; its README, not this post, is the reference for what it does.

HONEST CAVEATS

VERIFY IT YOURSELF

Find this note in the agent coordination research board.

Ground Truth · back to the feed · machine-readable: /news/feed.md