Sherwood Callaway has a habit of making bets that sound a little crazy at first. When he returned to Y Combinator this spring—his second tour through the storied accelerator—he arrived with a thesis that amounts to heresy in the observability world: the industry's foundational model is broken.
For years, companies monitoring their software systems have relied on what's known as the three-pillar approach. Logs, metrics, traces. Three separate data streams, three distinct pipelines, three different query languages. It's become gospel in the space—vendors build entire platforms around it, engineers spend hours configuring it, companies pay handsomely to maintain it.
Callaway thinks it's all unnecessary. His new venture, Sazabi, argues that logs alone can handle the job. Everything else? Let AI figure it out.
It's a contrarian stance, perhaps more so than Callaway himself expected when the company emerged from stealth in April. Because while every major player in observability is racing to sprinkle artificial intelligence onto their existing architectures, Sazabi is proposing something different entirely: tear down the architecture itself.
"Metrics and traces are just structured representations of events that already exist in your logs," Callaway told SiliconANGLE shortly after launch. The company's manifesto—published on March 16 with a deliberately provocative title—puts it even more bluntly: "Logs Are All You Need."
Whether that's visionary or just wishful thinking remains an open question.
Dismantling the Three Pillars
The technical argument goes like this. Logs, at their core, contain a complete record of what's happening in a system. When an API call fails, when a database query slows, when memory usage spikes—it's all there, timestamped and detailed. Metrics and traces, in this view, are simply pre-aggregated summaries of that raw information.
So why not keep the raw logs and reconstruct the summaries only when needed? That's Sazabi's bet. The platform uses AI to automatically summarize and organize log data, promising to eliminate the complexity—and substantial cost—of maintaining separate pipelines for each telemetry type.
Callaway, who previously built infrastructure and observability teams at Brex, says he watched engineers struggle with what he calls "observability sprawl." Configuring dashboards. Writing queries in three different languages. Managing three distinct data stores. The cognitive overhead alone was crushing; the infrastructure costs weren't trivial either.
Sazabi's answer is a conversational interface. Developers ask questions in plain English—"Why is checkout slow?"—and get back root causes, visualizations, suggested fixes. No query language required.
The autonomous alerts feature takes this further. According to the company's documentation, the system configures itself to detect issues: API error spikes, cloud cost anomalies, performance degradations. It correlates signals across commits, deployments, support tickets, all without manual setup. Sample alert payloads show surprisingly specific diagnoses—a misconfigured Redis connection pool here, a missing database index there, Lambda concurrency limits reached somewhere else. Each comes with recommended fixes.
It sounds almost too good. Which is precisely the gamble.
Building for the Machines
But the logs-only architecture isn't Sazabi's only departure from conventional wisdom. There's a second bet embedded in the product, one that might matter even more: the company is designing for AI agents, not just human developers.
The platform includes native integrations for Claude Code and Cursor. There's a command-line interface built for programmatic access. A Model Context Protocol server for IDE agents. Full REST and GraphQL APIs that make the system accessible to autonomous workflows.
"We're building for AI-native organizations," the company's YC profile states—a phrase that would have sounded like marketing speak two years ago but now describes actual engineering teams.
The CLI allows agents to query logs, tail live streams, instrument new data sources, all without switching contexts or waiting for human intervention. If AI is writing more of the code, the thinking goes, observability tools need to speak machine languages as fluently as human ones.
This design philosophy reflects Callaway's read on where software development is headed. Structured APIs and conversational interfaces instead of proprietary query languages. Automation by default instead of configuration. It's a vision of observability that assumes the developers using it might not be developers at all—at least not in the traditional sense.
Whether the market is ready for that remains unclear.
Angel Money and Ecosystem Bets
Sazabi's early backing tells its own story. The investor roster reads less like a typical seed round and more like a who's who of the AI developer tooling ecosystem.
Individual angels from Vercel, LangChain, Replit, Anthropic, Graphite, Daytona, Browserbase. Harrison Chase from LangChain. Matthew Lenhard and Andrew Qu from Vercel. Matt Palmer from Replit. These aren't random checks—they're strategic bets from people building the infrastructure Sazabi is positioning itself alongside.
Zypsy, an investor focused on design-forward products, published a piece around the launch framing their thesis around "observability for AI-generated code." The subtext: if AI is writing the software, the debugging tools need to evolve accordingly.
The startup participated in Y Combinator's most recent batch and took the accelerator's standard check. Directories show a $125,000 investment from YC, typical for the accelerator's standard terms, though no larger seed round has been announced. That's typical for this stage—prove the concept, land some early customers, then raise the real money.
But proof of concept is still very much in progress.
Crowded Territory, New Angles
The timing is both opportune and challenging. The observability market is in flux, with every major vendor pivoting toward AI in some form.
Recent months have seen a flurry of announcements. Grafana unveiled AI observability tools at its annual conference. IBM evolved its Instana platform to support AI agent and LLM monitoring. Datadog published research emphasizing investments in AI observability capabilities. Even M&A activity reflects the shift—Cisco announced plans to acquire Galileo Technologies to bolster Splunk's AI visibility.
Newer entrants are making similar moves. OpenObserve announced what it calls "Observability 3.0" with an autonomous AI SRE agent. Groundcover added agentic AI tracing support. The pattern is consistent: layer intelligence onto observability, make systems smarter about detecting and diagnosing issues.
What sets Sazabi apart—if it succeeds—is the architectural choice to start from scratch with logs only, building AI as the primary interface rather than retrofitting it onto existing three-pillar systems. The target market reflects this: early- to growth-stage technology companies that haven't yet accumulated the observability sprawl common at enterprises.
Companies, in other words, that can still make a clean break.
Proving Ground
As of now, Sazabi remains in pilot mode with a waitlist. No pricing has been disclosed. No general availability date announced. The company lists a team of 10-12 on its about page, including Callaway and co-founder Justin Ko, with whom Callaway previously built Opkit during an earlier YC batch.
Active hiring across engineering, marketing, operations, and sales suggests plans to scale beyond that core group. But product-market fit—that elusive milestone—is still being established. The YC announcement post from late March thanks early design partners without naming specific customers, a detail that speaks to just how early-stage this remains.
The fundamental question is whether AI can deliver what Sazabi promises: the ability to reconstruct the value of metrics and traces on demand, without the overhead of maintaining separate systems. It's a technical wager with significant architectural implications.
If Callaway is right, observability platforms could look dramatically different in a few years—leaner, more conversational, built for machines as much as humans. If he's wrong, Sazabi may find itself adding the very metrics and traces it currently argues against. The industry has seen bold architectural bets before; some pan out, many don't.
For now, though, the company is committed to its thesis. Logs, paired with sufficiently intelligent AI, really are all you need.
At least that's the bet.
