XMTP Labs

XMTP Labs builds secure, decentralized messaging infrastructure, powering end-to-end encrypted conversations in wallets and apps. The company has raised nearly $50 million, including a $20 million Series B co-led by Union Square Ventures and a16z crypto at a $300 million valuation. XMTP Labs also recently launched Convos, a secure consumer messenger with built-in agents that runs on its own protocol.

The challenge

XMTP Labs built a reputation on giving developers a secure, reliable protocol that lets wallets and apps like World (formerly Worldcoin) and Coinbase Wallet exchange encrypted messages without a centralized server. When the company decided to build Convos, a secure messenger with built-in agents, it expanded from a B2B infrastructure provider into a consumer product company for the first time.

As CTO, Nick Molnar carried a dual mandate: setting the technical direction for XMTP Labs’s core protocol and steering the team through its first consumer launch. “There was a six-week sprint to take this prototype of a new product and turn it into a real production service we could put into the App Store,” Nick says.

That timeline collided with a stack that was already complex. XMTP Labs runs services in Go, TypeScript, and Python, a client written in Rust, and mobile apps in Swift and Kotlin, all spread across AWS and Cloudflare. A bug could start in one service, surface as a symptom in another, and only make sense after the team traced the full chain of dependencies across tools and dashboards. Convos’s end-to-end encryption compounded the problem: the same privacy layers that make it trustworthy also complicate tracing an issue back to the right logs.

Convos changed the math entirely: more APIs, more ways to use the product, and AI agents inside conversations that could take actions on a user’s behalf, including some the team hadn’t fully anticipated. Six weeks wasn’t enough time to build alerting by painstakingly instrumenting each known behavior by hand. That coverage caught what the team already expected to go wrong, but offered no protection against blind spots — testers hit broken experiences before the team knew something was wrong, and engineers sometimes spent hours or days digging across services to understand what had actually happened.

Before real users hit production, Nick needed an observability layer built for the unknowns of an agent-heavy product. He chose Sazabi for its flexibility to surface issues on paths the team hadn’t instrumented yet.

How do you even figure out what to be looking for, in a world where you don’t know the full capabilities of your own product? We knew we needed to think about observability differently, because we were going to have blind spots.

The solution

XMTP Labs brought Sazabi online quickly to support Convos’s six-week sprint to the App Store. The team wired Sazabi into existing tools like Sentry and Datadog, began piping logs into Sazabi’s own storage, and had core setup live in about an hour, with one last service integrated shortly after.

From there, a clear day-to-day workflow took shape. Sazabi pushes alerts into a dedicated Slack channel for each environment. An engineer triages and asks follow-up questions in plain language, right in the thread. When the root cause isn’t obvious, Sazabi digs deeper — correlating errors with recent deploys, clustering similar failures together, and checking related services for the same signal. From there, the team either has Sazabi draft a first-pass fix or passes the issue, with full context, to their coding agent to implement in the normal development workflow.

One early example showed exactly what that coverage catches. Convos’s agent was silently failing to deliver images above a certain size, a rare edge case that internal testing hadn’t surfaced. Sazabi flagged it directly from the logs, turning a silent failure into a clear, traceable issue. “That kind of thing is very easy to miss because it only happens a handful of times a day. But it can hurt trust in the product, and we want to fix those bugs,” Nick says.

What’s surprised him most is how little the team has had to teach Sazabi. Rather than relying on hand-built rules for known failure modes, it ingests and stores logs, then learns the system’s normal behavior. The platform spots patterns that aren’t covered, flagging gaps and suggesting new alerts with almost no context from the team. New blind spots are covered automatically as Convos evolves, closing the unknowns and uninstrumented paths other tools miss.

That precision removes alert fatigue: repeated alerts fold into a single issue, so an engineer sees one thing to act on instead of five versions of the same problem. More than 80 percent of Sazabi’s alerts correctly require a human response, a sharp break from the noisy, low-signal alerts that used to eat into the team’s time.

Sazabi’s reach has extended beyond engineering, too. Connected to XMTP Labs’s product analytics, it lets the go-to-market team ask questions directly in Slack, without waiting on an engineer to write the query. Non-engineers now investigate on their own, reducing ad-hoc data pulls and freeing engineers to stay focused on the roadmap.

We’re catching things we wouldn’t have known about until someone complained. That’s the biggest speedup Sazabi has given us.

Results

With Sazabi in place, XMTP Labs moved from reactive monitoring of a few known protocol paths to proactive coverage across a fast-changing consumer product. Detection and root-cause analysis are quicker, and subtle bugs that once took days to surface are now caught in minutes, before they show up as broken user experiences.

• 5 minutes to flag a major incident with enough context to start fixing immediately
• 20 minutes to ship a fix for a critical issue once Sazabi raises an alert
• 80%+ of alerts correctly require a human response, reducing noisy, low-signal alerts

Looking ahead, Sazabi gives the team the confidence to keep shipping new agent capabilities without trading away reliability. Writing solid logging takes far less expertise than instrumenting metrics and traces, enabling more contributors to participate in building Convos.

As we keep shipping, Sazabi’s logs-only approach means people can do the right thing the first time, without needing a lot of fine-tuning and expertise.