I moderated a panel last week on infrastructure for AI agents in messaging, and one observation kept coming up: the chatbot era taught us almost nothing useful about what’s coming next.

Those early bots — the “type 1 for billing, type 2 for support” kind — were basically IVR menus in a text box. Rigid decision trees. No real state management. No tool orchestration. No fallback beyond “I didn’t understand that, please try again.” You know the drill.

What’s emerging now is fundamentally different. We’re talking about AI agents that live inside messaging threads and autonomously execute multi-step tasks: checking inventory, initiating refunds, scheduling appointments, making API calls to backend systems — all while maintaining conversational context across turns and sessions.

The infrastructure layer to support that? Barely exists. It’s being built right now, in real time, and the design decisions happening today will shape agentic messaging for the next decade.

Why RCS Is the Channel #

I’m biased here — I work in this space — but RCS has architectural advantages that make it uniquely suited for agentic AI.

Start with the protocol itself. RCS Business Messaging supports rich cards for displaying structured data: product catalogs, order summaries, booking confirmations. It has suggested actions — tappable buttons that trigger specific agent behaviors. Verified sender identity means users know they’re talking to the real brand, not some phishing attempt. And GSMA’s MaaP (Messaging as a Platform) framework provides a standardized way to build conversational experiences on top of the messaging layer.

Compare that to SMS: 160-character plaintext and zero structured interaction. Or WhatsApp Business, which has rich features but operates inside Meta’s walled garden with all their API limitations. RCS sits at the carrier layer with operator-level reach and an open standard for extensibility.

Apple’s adoption of RCS (starting with iOS 18.1) changed the calculus entirely. Before that, RCS was basically an Android story. Now it’s cross-platform, and the addressable market for RCS-based AI agents expanded by roughly a billion devices overnight.

The market numbers reflect this shift. RCS messaging revenue is projected to hit $45 billion by 2033, driven significantly by AI chatbot integrations, verified messaging, and cloud platform adoption. That’s not a niche bet.

The Infrastructure Stack #

When I say “infrastructure for AI agents in messaging,” I’m not talking about the LLM. The model is the easy part — relatively speaking. I’m talking about everything around it that makes agentic behavior possible in a messaging context.

Message routing and orchestration. An AI agent in messaging needs to receive messages, determine intent, orchestrate tool calls, and compose responses — all within carrier-imposed latency constraints. RCS Business Messaging has delivery requirements; if your agent takes fifteen seconds to respond because it’s chaining three API calls, the user experience falls apart. You need an orchestration layer that can parallelize tool calls, manage timeouts, and return partial results when full resolution takes too long.

State management across sessions. This is where most current solutions fall apart. A user starts a conversation about returning a product on Monday, gets distracted, picks it up Wednesday. The agent needs to remember the context: what product, what’s the order number, what was the last thing discussed. Messaging platforms don’t maintain this state for you; your infrastructure has to. And it needs to handle concurrent conversations (same user, same agent, two devices), session expiration, and GDPR-mandated data retention limits.

Channel fallback. Not everyone has RCS. Your agent needs graceful degradation. If the user’s device doesn’t support RCS, fall back to SMS — but an SMS fallback of a rich card experience requires reformatting the interaction entirely. Drop the carousels, linearize the choices, convert suggested actions to numbered text options. This isn’t trivial; it’s essentially maintaining two conversation UIs for the same agent logic.

Observability. When an AI agent handles ten thousand conversations simultaneously, how do you know it’s performing well? You need conversation-level metrics (resolution rate, escalation rate, user satisfaction), system-level metrics (latency, error rates, tool call success rates), and content-level monitoring (is the agent saying anything it shouldn’t?). The observability infrastructure for agentic messaging is still immature. Most teams I talk to are stitching together logging, APM tools, and custom dashboards.

Compliance and audit trails. Financial services, healthcare, government — regulated industries require audit trails for every decision an AI agent makes on behalf of the organization. If your agent approves a refund, there needs to be a record of why, based on what information, and whether it followed policy. This audit infrastructure has to be built into the agent architecture from the start, not bolted on after launch.

Who’s Building This #

The ecosystem is coalescing around a few approaches.

Infobip integrated Google Vertex AI directly into their RCS infrastructure, enabling AI-powered conversational agents that operate natively within RCS Business Messaging. That’s significant because Infobip has carrier-level reach across hundreds of operators. When they wire up Vertex AI to RCS, it’s not a demo — it’s production infrastructure that can reach billions of devices.

Linq raised a $20 million Series A specifically to build the infrastructure layer for AI messaging agents. They’re already powering over 30 million messages per month. What’s interesting about their approach is the focus on the plumbing: message routing, state management, analytics. They’re not building the AI agent itself; they’re building the platform that AI agents run on. That’s a bet that the infrastructure layer is where the value accrues — and I think they’re right.

Twilio’s RCS platform went GA in August, and their Programmable Messaging APIs provide another foundation for agentic messaging. Twilio’s advantage is developer mindshare — millions of developers already know how to build on Twilio APIs.

McKinsey reports that 60% of companies are exploring AI agents for customer engagement, with 85% expected to implement by end of 2025. Those numbers suggest we’re past the experimentation phase and into infrastructure buildout. The companies I see succeeding are the ones treating this as a platform problem, not a chatbot problem.

The Hard Problems #

Three challenges came up repeatedly during the panel, and none have clean solutions yet.

Context window vs. conversation history. Messaging conversations can span days or weeks. The accumulated context — previous orders, preferences expressed three conversations ago, complaint history — exceeds what you can stuff into an LLM context window. You need a retrieval layer that selectively surfaces relevant history without overwhelming the model. This is essentially RAG applied to conversation memory, and doing it well across thousands of concurrent sessions is an infrastructure challenge, not an AI challenge.

Rate limits and carrier compliance. Carriers impose rate limits on RCS Business Messaging. If your AI agent suddenly needs to send 50 messages to 50 users in the same second (say, order status updates triggered by a batch event), you need queuing, rate limiting, and retry infrastructure that respects carrier constraints without dropping messages. Get this wrong and carriers throttle or suspend your agent.

Trust and verification. When an AI agent executes a transaction inside a messaging thread, both parties need assurance. The user needs to trust that the agent is authorized to act on the brand’s behalf. The brand needs to trust that the agent’s actions are within policy. RCS’s verified sender framework handles the first part. The second part — policy enforcement for autonomous agent actions — requires its own infrastructure layer.

Where This Is Heading #

The messaging channel is becoming the execution layer for AI agents, not just the communication layer. An agent that can check your flight status, rebook you on an alternative, charge the fare difference to your card, and send you the new boarding pass — all within the messaging thread, without ever redirecting to an app or website — that’s the direction we’re headed.

Building the infrastructure to support that reliably, at scale, across carriers and devices, with proper compliance and observability — that’s the actual hard work. The LLM is the brain; the infrastructure is everything else: the nervous system, the skeletal structure, the immune system.

We’re in the early innings of getting this right. The companies investing in the infrastructure layer now are building the rails that agentic messaging will run on for years.