Ably

Airbnb's engineers are shipping 80% more features than a year ago. Here are a few things they're building with that momentum: 1/ A CUSTOM-BUILT SUPPORT AGENT Yashar Mehdad, Airbnb's Principal AI Architect, has described the AI assistant as "a custom built AI agent product developed with multiple LLMs and components." Ted Burke owns it as a product, with CS AI engineering under Claire (Na) Cheng and Kelvin Xiong. It works from your reservation details and can suggest the next step, like cancelling a trip. By February, it was resolving about a third of issues without needing an agent. It now chats in 50+ languages and voice support is next. 2/ HANDOFF TO A PERSON IN THE SAME CHAT When the assistant can't help, guests can carry on with a person in the same chat or switch to a phone call. Omar Siddiqui leads the human-in-the-loop platform behind Airbnb's support AI. If my flight is delayed and I've already explained it to the assistant, I want the person who picks up to carry on from there. 3/ AN INTERNAL AI TOOLING PLATFORM CTO Ahmad Al-Dahle said last month: "Our development teams are shipping roughly 80% more features than a year ago." Airbnb engineers use an internal AI tooling platform to write software and create remote AI agents. Kjell Bronder leads AI product strategy across the guest journey, including search and trip planning. Last week's autumn update, which put AI into search, comparisons and support is that speed reaching guests. There's one part I'm keen to see Airbnb handle. It builds much of this itself: its own support agent, a human-in-the-loop platform behind it, and AI tooling for its engineers. In my experience, once AI reaches search, comparisons, support and voice, the hard part is keeping each of those experiences responsive and in sync for the guest. What I've seen work is building that once, at the platform level, so each product team doesn't have to solve it alone. That's where teams like Sagar N. s Communications Platform come in. It also frees more of the 80% for features guests see. Airbnb already uses AI to help guests when travel plans change. That's the moment a guest remembers, when the plan falls apart and one conversation gets them back on track.

Choosing WebSockets for AI streaming is the easy part. Running them reliably is where the work starts. A WebSocket gives users a way to interrupt or approve an agent mid-stream. It doesn’t give you session recovery, fan-out across devices, or delivery guarantees. Your team has to provide those – or use a managed layer such as Ably AI Transport. When does running it yourself make sense? Madeleine Quinn walks through the technical and cost trade-offs: https://lnkd.in/eYeS8qsy

CTO: "We need live features." Engineer: "Easy. It's just WebSockets. Give me 3 engineers and 1 month." 3 months later: "Finally shipped! …but reconnects are dropping messages." 6 months later: "Messages are arriving out of order now we're on multiple servers." 9 months later: "Turns out you can't autoscale WebSockets like HTTP. Also, a region went down." 12 months later: "AI responses are cutting off mid-stream." 18 months later: "Who's turn is it to be on call this weekend?" CTO: "And what happened to the roadmap?" Building realtime is easy in theory. Running it forever is the reality.

A regional failover can turn one outage into two. Here are three assumptions that cause it: 𝗔𝘀𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻 #𝟭: 𝗘𝗮𝗰𝗵 𝗿𝗲𝗴𝗶𝗼𝗻 𝗼𝗻𝗹𝘆 𝗻𝗲𝗲𝗱𝘀 𝗰𝗮𝗽𝗮𝗰𝗶𝘁𝘆 𝗳𝗼𝗿 𝗶𝘁𝘀 𝗼𝘄𝗻 𝘁𝗿𝗮𝗳𝗳𝗶𝗰 Teams often scope multi-region as the same deployment running somewhere else. Then a region fails with a million connections on it, and those clients reconnect at once into a region with no headroom for them. Plan capacity for the failover case: - The extra connections each region absorbs when a neighbour goes down - The reconnect spike in the first few seconds, well above steady state 𝗔𝘀𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻 #𝟮: 𝗔 𝗿𝗲𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗲𝗱 𝗰𝗹𝗶𝗲𝗻𝘁 𝗵𝗮𝘀 𝗿𝗲𝗰𝗼𝘃𝗲𝗿𝗲𝗱 A client that reconnects still has to resume from the last message it saw, with nothing lost or delivered twice. A publisher in Frankfurt still has to reach a subscriber who just landed in Virginia, so messages need a path across regions as well as within them. 𝗔𝘀𝘀𝘂𝗺𝗽𝘁𝗶𝗼𝗻 #𝟯: 𝗧𝗵𝗲 𝗻𝗲𝗮𝗿𝗲𝘀𝘁 𝗵𝗲𝗮𝗹𝘁𝗵𝘆 𝗿𝗲𝗴𝗶𝗼𝗻 𝗰𝗮𝗻 𝘀𝗲𝗿𝘃𝗲 𝘁𝗵𝗲 𝘂𝘀𝗲𝗿 If you persist message history, residency rules decide which regions can serve a user, and the nearest healthy one may be off-limits. Connection routing and data storage can live in different places, but only if you design for it from the start. Dependencies count too. If auth in the failover region is down, a user can land somewhere healthy and still not get in. 𝗪𝗵𝗲𝗿𝗲 𝗜'𝗱 𝘀𝘁𝗮𝗿𝘁 Take one user session and trace it through a regional outage: - Where does the connection land? - What state does it need to resume? - Which regions is it allowed into? - What happens if auth there is down too? Then split what you find into two lists: what your application needs to own and what your realtime infrastructure should handle for you.

Wix wants to run the small business, not just build its website. Here is why: Most website builders are adding AI the same way. They put an assistant in the editor, ship a chatbot template and call it an AI strategy. It rarely changes how the business owner actually runs their day. The agent only lives inside one tool, doing one job, in isolation. Wix is going much further. It has built multiple agents that coordinate with each other and keep the owner updated in realtime, not a single bot added to an editor. 1/ A STANDALONE MULTI-AGENT SYSTEM On 11 August, Wix launched Symphony, a multi-agent system built for SMBs. Maestro is its central orchestrator agent. Its agents "can work across approved workflows on the owner's behalf", on the devices, channels, voice interfaces and AI platforms they already use. 2/ AGENTS WHERE THE CUSTOMER ALREADY IS Wix was the first CMS to sign the Agentic Commerce Protocol (ACP). Aria, Juno, Kleo and Smart Chat each cover a different part of the business. Since 9 September, Wix has also been in OpenAI's new SMB directory. An owner can build and manage a Harmony site from inside ChatGPT. 3/ INFRASTRUCTURE BUILT FOR AGENTS When Anthropic came to Wix to test its commerce agent, Nir Zohar summed it up: "That's a good sign our infrastructure is ready for what's coming." My prediction is that Symphony's real strength will be how well Wix's agent teams work together. A business owner might approve something by voice, check on it in ChatGPT, then answer a customer in Smart Chat. Product, engineering, support and partnerships each own a piece of that. It has to feel like one experience. Asaf Yonay wrote this month: "The ability to build something was never a sufficient reason to build it." With that many channels to cover, Wix picks its battles. It signed up to ACP instead of inventing its own standard. Wix Engineering also published an arXiv paper on Helpmate, its Customer Care AI agent, tested across 756K+ production user messages. Wix is writing its agent playbook down, in public, while the platform is still young. That's the team I'd be watching over the next year. Aviran Mordo 🪐, Even-Haim Yaniv, have I missed anything important?

I asked Matthew Hammond, Head of Infrastructure at Ably, what he'd do if our user base grew 10x overnight and how he'd keep the infrastructure from breaking the agent-user experience? Here's what he said: "Surviving 10x overnight starts with ensuring there is headroom and a proven ability to add infrastructure capacity quickly, but less obvious risks sit in the ecosystem that exists around all platforms. Will deployments still be safe and fast, or slow down just as you need to roll out bug fixes, new features or performance efficiencies? Will your monitoring and observability systems cope with the surge in events, logs and metric cardinality, or go down just as you need assurance that the system is working as expected? The agent-user experience only holds up if everything that supports it scales with it. Infrastructure teams have got to ensure that all supporting systems and operations remain viable too as a platform scales." There's an uncomfortable timing problem here. The moment you most need your monitoring and deployment systems is also when they're under the most pressure. It makes me wonder how many teams think of those systems when they ask themselves if they're ready for 10x. We're running new research at Ably into how engineering leaders protect the agent-user experience as they scale. I started close to home with Matt, but I'm keen to hear how teams outside Ably are handling it, especially what they only found out in production. People I'd like to invite: Asaf Yonay, GM of AI-Native Transformation & AI Platforms at Wix Diego Comas, Senior Director, Platform & Operations at Sourcegraph Leonid Belkind, CTO and co-founder at Torq If you have a point of view, comment or send me a message. I'd love to include your voice.

Talkdesk shipped 4 AI launches in 6 months. Three things its engineers are doing caught my eye:   1/ BRINGING CUSTOMER CONTEXT INTO BRANCHES AND STORES Since March, Talkdesk has launched the Customer Experience Automation (CXA) Operations Center, Agent Builder, Talkdesk for Financial Centers and Talkdesk for Stores. Together, these cover how enterprises build agents, manage them alongside people, and bring AI assistance into more customer interactions. The branch and store products are particularly interesting. They give employees access to earlier interactions and AI guidance while they’re speaking to a customer in person. If someone has already explained their problem to the contact center, the store associate has something to work from (major CX win!). 2/ MAKING AI CONVERSATIONS VISIBLE WHILE THEY HAPPEN Munil Shah, Talkdesk’s Chief Product, Technology and Customer Officer, has described the industry gap saying teams are “deploying AI agents without a rigorous way to evaluate behavior, reliability, tool usage.” Talkdesk’s documentation gets specific about the supervision. People can monitor running AI sessions as the conversation unfolds. On voice calls, an AI agent can escalate to a person and pass along the data it has collected. Anyone who’s had to explain the same problem to three different people will appreciate why that matters. The handoff is part of the customer experience, and it deserves the same engineering attention as the agent (another major CX win!). 3/ APPLYING AI TO THE INFRASTRUCTURE WORK TOO Mudit Mathur was recently leading Talkdesk’s 50+ engineer Core Platform group, whose work included AI-native tools and self-service AI agents for infrastructure management. Ana Afonso runs two teams in the Core Platform Cluster building distributed systems. Alongside them, Filipe Lima is VP of Engineering, Platform, and Yunjing Ma leads AI Engineering, Research & Applied AI. The investment reaches into the day-to-day work of running the platform as well as the products customers use. Prashanth Padmanabhan, VP of AI Product Management for Forward Deployed, puts it plainly “Enterprise customers don’t buy AI – they buy outcomes.” I’d apply that test inside Talkdesk, too. If the platform team’s AI agents take infrastructure tasks off engineers’ plates that gives them more time to build what comes next. Talkdesk for Stores makes the customer outcome easy to picture. A customer walks in, the associate can see what they’ve already discussed, and they can get on with helping them.

How many WebSocket fixes does it take before you’re maintaining your own realtime platform? Reconnection, cancellation, multi-device delivery, crash detection. The work adds up. Full article in the comments. #AIStreaming #WebSockets #Realtime