Here's a problem you don't see coming: an AI agent circles the same decision tree fourteen, maybe fifteen times, burning through API credits while your users quietly lose patience. No error message. No crash. Just a conversation slowly spiraling nowhere, and by the time anyone notices, the damage is done.
Sentrial, a two-person team fresh out of Y Combinator's latest batch, thinks it has an answer. The startup launched its production monitoring tool on February 7 with a specific pitch—catch the ways AI agents fail that traditional observability platforms never see. Infinite loops. Hallucinations. Tool misuse. User frustration building in real time. Not three hours later when some dashboard finally flags an anomaly. Right now, as it's happening.
Whether that's enough to compete with the giants already circling this space is another matter entirely.
The Rush Is On
Timing here isn't accidental. The past few weeks have seen a land grab in agent monitoring, with incumbents moving fast to claim territory. New Relic unveiled its Agentic Platform in late February (the company says it was February 24, though exact timelines in this fast-moving sector can blur). Dynatrace rolled out what it's calling "Dynatrace Intelligence" in late January, positioning agentic AI as central to enterprise observability going forward. ClickHouse, the database company, acquired Langfuse—an open-source LLM observability project—on January 16.
The category is consolidating before it's even fully formed.
Sentrial's founders, Neel Sharma and Anay Shukla, graduated from UC Berkeley's computer science program and came at the problem from opposite ends. Sharma spent time working on agentic optimization at Sense. Shukla deployed agents in production at Accenture, navigating the messy reality of autonomous systems that don't behave like traditional software. Their bet: enterprise observability vendors, however fast they're moving, will struggle to truly understand how agents break.
"APM tools show you latency and error rates," Sharma noted in a developer-focused post. "They don't tell you that your agent asked for the same billing information twice because it lost context mid-conversation."
Four Flavors of Failure
Sentrial's approach centers on detecting four specific failure modes that don't map neatly onto the world of HTTP status codes and response times.
Infinite loops, where an agent retraces identical steps without making progress. Hallucinations—a term that's become ubiquitous in AI circles—where model outputs drift from factual grounding into something plausible-sounding but wrong. Tool misuse, when an agent calls the wrong API or fumbles the parameters. And user frustration, which the system attempts to infer from conversational patterns that signal confusion or disengagement.
The diagnostic layer runs at runtime, analyzing conversation flow, model outputs, and tool interactions to surface not just that something broke, but why. According to documentation for the company's PyPI package, Sentrial can even recommend code changes and generate GitHub pull requests automatically.
Whether engineering teams actually merge AI-generated PRs is, of course, an open question. But perhaps the more valuable piece is the diagnosis itself—identifying that an agent is stuck, not because of network latency, but because it's contextually adrift.
A Narrow SDK in a Crowded Landscape
The technical implementation is deliberately lean. Sentrial's SDK—which reached version 0.5.2 on February 28—supports OpenAI, Anthropic, and Google's Gemini models. It hooks into popular frameworks like LangChain (both legacy and current versions), CrewAI, and AutoGen through wrappers. Teams already instrumented with OpenTelemetry can route traces to Sentrial's backend with a setup function.
The company's developer site promises "drop-in observability for LangChain" via a single callback handler. Marketing materials claim "thousands of developers" are using it, though no customer logos have appeared publicly, and there's no pricing information listed—just demo requests and waitlist sign-ups. That opacity is typical for an early-stage YC company barely six weeks past launch, but it also makes independent validation difficult.
LangSmith, LangChain's own observability product, has published guidance on why monitoring agents differs fundamentally from traditional application performance monitoring. Sentrial's positioning follows a similar logic: agents aren't microservices. They make probabilistic decisions, chain together tool calls, and fail in ways that resist clean categorization.
David vs. Several Goliaths
Which brings us to the central tension. Can a two-person startup carve out defensible ground while Dynatrace, New Relic, and Oracle (which recently integrated Arize Phoenix with its Open Agent Spec) build out their own agent monitoring capabilities? The enterprise players have installed customer bases, compliance certifications, procurement relationships stretching back years. Sentrial has a focused product and a Python package.
The advantage, if there is one, comes down to velocity and architectural assumptions. Incumbent platforms are retrofitting agent monitoring onto observability stacks built for a different era. Sentrial designed from the premise that agents fail differently. If agent adoption accelerates faster than enterprise vendors can retool decades-old telemetry pipelines, that head start might matter.
Security vendors are entering from yet another angle. Operant AI launched Agent Protector on February 5, framing real-time behavioral threat detection as adjacent to observability. The boundaries between monitoring, security, and guardrails are getting messier. Sentrial's focus stays narrower: diagnose what broke, explain why, help fix it.
The Bet Beneath the Product
For engineering leaders already running agents in production—and more are every month—the question isn't whether to monitor. That's settled. The choice is whether to wait for an existing observability vendor to ship agent-aware features, or adopt a purpose-built tool now and accept the integration overhead.
Sentrial's wager is that the gap between "traditional observability with AI bolt-ons" and "designed for agents from scratch" is meaningful enough to justify the friction. Six weeks into the company's life, with no disclosed funding beyond Y Combinator's standard batch investment and a two-person team, the product will likely evolve significantly. Features could expand. Pricing models will emerge. Customer traction will either materialize or it won't.
But the underlying thesis—that autonomous agents fail in ways conventional monitoring systems aren't built to catch—is already playing out across the industry. Whether Sentrial captures that shift or gets absorbed into a larger platform's roadmap depends on how quickly they can move while the category is still taking shape. And whether enterprises, once they start feeling the pain of agent failures slipping through their existing tools, decide a startup's focused solution beats waiting for the incumbents to catch up.
For now, the infinite loop problem is real. The market's still deciding who gets to solve it.
