Photon Spectrum: Bring AI Agents to iMessage, WhatsApp and Telegram

What Is Photon Spectrum?

Spectrum is Photon's answer to a channel problem: a user might message you in iMessage today and WhatsApp tomorrow, and each channel normally means a separate integration with its own authentication, events and message formats. The Spectrum documentation puts one agent server behind all of them, with a provider for each interface:

Photon's homepage also lists SMS, RCS, Discord and Slack. SMS and RCS appear there as fallbacks for iMessage delivery, the docs treat Discord as a custom-provider target, and Slack is not among the documented providers, so check the provider pages before you plan a channel.

How Does a Spectrum Agent Server Work?

Your Spectrum server owns the product behavior, meaning routing, tools, memory, handoff, safety and analytics. Each provider owns the interface work of connecting to the platform, receiving events and sending messages. Every provider feeds one message stream, so the agent loop stays the same when you add a channel, as in this example from the docs:

import { Spectrum } from "spectrum-ts";
import { imessage, terminal } from "spectrum-ts/providers";

const app = await Spectrum({
  projectId: process.env.PROJECT_ID!,
  projectSecret: process.env.PROJECT_SECRET!,
  providers: [
    imessage.config(),
    terminal.config(),
  ],
});

for await (const [space] of app.messages) {
  await space.send("How can I help?");
}

You install it with npm install spectrum-ts, and it requires TypeScript 5 or later. The code above works with four primitives:

• Message - an incoming piece of content, whether text, attachments or structured data, from any platform.
• Space - a conversation context such as a DM, a group chat or a terminal session. You send messages into a space.
• User - a participant on a platform, identified by a platform-specific ID.
• Platform provider - the adapter that translates one platform's protocol into Spectrum's unified interface.

Spectrum handles channels, not reasoning, so it sits beside whichever framework builds the agent itself; Photon's blog shows it with Mastra and Convex agents. The agentic AI frameworks guide compares the options for that layer.

What Can a Photon Agent Do Inside iMessage?

Photon's documentation describes iMessage as the channel where Spectrum is most mature.

Photon's homepage lists the native iMessage APIs an agent can use:

• Send and receive messages with markdown and bubble or message effects.
• Share attachments, including files, media and stacked images.
• Accept location sharing and subscribe to real-time location updates.
• Build and deliver generative iMessage mini apps inside the conversation.
• Set the agent's identity with profile names and photos.
• Create and manage group chats, including names, avatars and participants.

Those features matter most when the agent acts on real accounts. Flip, an AI financial assistant, runs on iMessage through Photon: users text it to pay people back, move money, split bills in group chats and track a portfolio, and Flip waits for a yes before any action that touches money. Photon reports that Flip circulates over $57 million in connected volume.

What Makes Messaging Agents Hard to Get Right?

Photon's inbound pipeline guide opens with how people text: in bursts such as "hey", "wait", "actually" and then the real question. An agent that answers each message sends overlapping replies and never sees the question, so the guide debounces for a few seconds and handles the burst as one turn.

Each item below is a failure a test suite for a messaging agent should cover:

• Message bursts - send several short messages in a row and confirm the agent answers once, to the whole burst.
• Cancelled jobs - the inbound guide keeps messages in a queue table until the handler reads them, and carries drained messages forward if a job is cancelled mid-generation, so nothing a user typed is dropped.
• Duplicate replies on retry - a worker that crashes halfway through a multi-message reply resends the first messages on retry. The recovery and state guide prevents this with stable client GUIDs and a persisted startIndex cursor.
• Memory that leaks between people - one agent talks to many users, and the recovery guide scopes working memory per person and history per thread, because telling one user about another user's plans is the bug it warns about.
• Voice call security - Spectrum Voice carries call audio over RTP, not SRTP, so audio on the SIP side of a call is not encrypted, while SIP signaling can use TLS.
• Channel fallback - Photon's API can route a message through iMessage with Telegram as a fallback, so test that a reply still reads correctly in the channel it actually lands in.

The terminal provider makes most of these reproducible on a laptop: Photon describes it as a drop-in test harness with multiple conversations, typing indicators, reactions, threaded replies and attachments. Pair it with a chatbot testing plan, then add the channel-specific cases above.

How Do You Test a Photon Agent Before It Ships?

A Photon agent can fail in the conversation, by saying the wrong thing, or in the world, by doing the wrong thing on a real account. With TestMu AI, Agent Testing grades the conversation and Agent Assurance checks what the agent actually did, so each risk below lands with the tool built to catch it:

npm install -g @testmuai/rook

Getting Started With Photon and TestMu AI

Build the agent against Spectrum's terminal provider first, so you can replay bursts, group threads and crashed retries locally before you connect a live iMessage line. Then list every action the agent can take on a real account; that list becomes your Agent Assurance criteria.

Generate conversation scenarios in Agent Testing from your PRD, run rook against staging so each action is checked against what changed, and keep both suites in CI so every prompt or model change is tested before it reaches a user's Messages app.