In a cramped Slack channel somewhere, an AI agent is doing what most engineers dread: investigating a production incident at 3 a.m. It's tailing logs, correlating errors, opening pull requests with fixes. No dashboards required. No human operator squinting at metrics. Just logs—and a bet that they contain everything you actually need.
That bet belongs to Sherwood Callaway, whose startup Sazabi emerged from Y Combinator's Spring 2026 batch with what might charitably be called a controversial thesis. The observability industry has convinced companies they need three things: metrics, traces, and logs. Datadog built an empire on that trinity. Callaway thinks it's bloat.
Sazabi ingests only logs, then leans on AI agents to reconstruct the rest. It's an architecture that, depending on your vantage point, either strips observability down to its essence or throws away half the toolkit just as systems are getting more complex.
The timing is deliberate. AI workloads are flooding production environments, and the old playbook is showing its age. A May 2026 survey from groundcover found that AI systems now account for up to half of observability spending at some companies—a cost spiral that has executives asking hard questions about instrumentation overhead. Callaway, who previously built infrastructure and observability teams at Brex, believes traditional tools were designed for a world where humans read dashboards, not one where agents ship code autonomously.
"AI lets us compress and summarize at ingestion," he told SiliconANGLE in April, describing a storage strategy that sidesteps the need for separate telemetry pipelines. If logs capture sufficient context—and if you can process them intelligently—why pay to maintain the rest?
A Manifesto Against Dashboards
Sazabi's pitch borders on austere. Connect your GitHub repository, install the Slack app, send OpenTelemetry-compatible logs. The platform handles the investigation work: reconstructing metrics and traces from log streams, running automated root-cause analysis, surfacing issues through conversational threads instead of dashboard sprawl.
The company laid out its philosophy in a March 2026 manifesto with headings that read like provocations: "Less is More," "Logs Are All You Need," "Monitoring Is Dead." It's the kind of declarative positioning that works in a Y Combinator pitch deck, though whether it survives contact with enterprise procurement is another question.
Early usage numbers, disclosed on Sazabi's YC Launch page in June, offer a glimpse of traction—or at least experimentation. In the month prior, the company onboarded 35 teams, ingested 2 terabytes of logs, ran 8,000 background investigations, and opened 200 pull requests with suggested fixes. The platform remains in closed alpha, which means the real test—whether this scales beyond early adopters with high tolerance for rough edges—is still ahead.
Built for Machines, Not People
Where Sazabi diverges most sharply from incumbents is in how it exposes data. Every feature is programmatically accessible: a Model Context Protocol server, REST and GraphQL APIs, a command-line interface. Engineers can tail logs or investigate errors inside Cursor or Claude Code without toggling to a browser. The company frames this as "agent-friendly" design, built for machine consumption rather than human eyeballs.
The autonomous alert system learns baselines, correlates changes to incidents, then posts in Slack with impact analysis, suspected root causes, and recommended actions. Alert categories span silent failures, error spikes, cost anomalies, slow queries—the usual suspects, but delivered as conversational summaries rather than time-series graphs. It's a workflow designed for teams that want observability to feel less like archaeology. Whether it actually does remains to be seen.
Integrations cover more than 35 cloud and hosting services: Vercel, AWS, GCP, Temporal, Cloudflare, Neon, Supabase. Sazabi accepts logs via OpenTelemetry, Syslog, and JSON—standard formats, funneled through a decidedly non-standard backend.
The Infrastructure Insiders Are Interested
Sazabi's angel investor list reads like a reunion of the developer infrastructure class of 2020-something. Harrison Chase from LangChain, Matt Biilmann from Netlify, operators from Vercel, Replit, Graphite, Daytona, Browserbase—all named on the company's YC profile. As of early April, Sazabi was raising a seed round, though the amount hasn't been disclosed.
Callaway's pedigree helps explain the interest. He previously co-founded Opkit, a YC Summer 2021 startup that automated insurance verification for telehealth, and later worked as a tech lead at 11x before launching Sazabi. His Opkit co-founder, Justin Ko, was also a Brex alum. The Sazabi team now numbers 10, according to the company's profile.
In an industry where credibility often hinges on who you've worked with, that roster matters—perhaps more than the product itself at this stage.
A Market Already Pivoting
Sazabi isn't the only company betting that AI changes the observability game. Honeycomb launched "Agent Observability" in May 2026 to bring visibility to agentic workflows. OpenObserve announced a $10 million Series A in late April, positioning itself as "AI-native." New Relic debuted an AI agent platform and OpenTelemetry tooling in February. Even Datadog—whose State of AI Engineering report dropped in April 2026—acknowledged that AI introduces new operational complexity.
The competitive landscape on Sazabi's YC Launch page is revealing. The company contrasts itself not just with legacy players like Datadog, Grafana, and Sentry, but also with "AI SRE overlays" that bolt agents onto existing tools, DIY internal assistants, and LLM-specific platforms like Arize and Braintrust. Sazabi's claim is vertical integration: one system handling the interface, agent logic, and storage, rather than stitching together layers.
It's a clean narrative. Whether it's a durable one depends on questions that won't be answered in alpha. Can AI reliably extract signal from unstructured log streams at scale? Will developers trust conversational debugging over the visual dashboards they've spent years learning to read? Does a logs-only architecture hold up under the weight of a distributed system with hundreds of services?
Logs, Agents, and the Bet on Simplicity
For now, Sazabi is a closed alpha with a waitlist and a thesis. The thesis holds that observability built for the AI era shouldn't resemble observability built for the dashboard era—that the three-pillar model is a relic of human-first operations, and logs processed by intelligent agents can do the job with less overhead.
It's an elegant idea, and elegance has a way of attracting true believers in Silicon Valley. But production systems have a way of resisting elegance. The strain of AI workloads may indeed expose mismatches in traditional tooling. Whether it vindicates Callaway's minimalist bet or simply creates demand for yet another category of instrumentation is the kind of question that only time—and a few late-night incidents—will answer.
