Two engineers from Y Combinator have a pitch that sounds either wildly ambitious or perfectly timed, depending on how much faith you place in autonomous systems: What if your monitoring tools didn't just alert you to problems, but fixed them?
Superlog—a startup so lean it lists exactly two people on its team page—launched in mid-May with software that claims to do precisely that. Install it, let it watch your code, and when something breaks, it investigates the failure, drafts a fix, and submits a pull request. No dashboard-staring required.
The product surfaced on Launch YC around May 12–13 and hit Show HN on May 19, where it collected somewhere between 72 and 92 upvotes along with a cascade of pointed questions. Developers wanted to know: dry-run mode? Which files does this thing actually touch? And perhaps most urgently—do I really trust an agent to merge code into my main branch?
Fair concerns, all of them. But the concept taps into something real: a fatigue with observability platforms that generate noise without resolution. Datadog and Sentry will tell you what's broken. Superlog wants to go ahead and fix it.
How the System Actually Works
The onboarding starts with a wizard that scans your repository and layers in OpenTelemetry instrumentation—logs, metrics, traces—without manual setup. Most observability stacks require engineers to plant these hooks by hand and keep them updated as code evolves. Superlog's wizard runs daily, adjusting instrumentation as your codebase shifts. That's table stakes for what comes next.
When errors surface, the platform groups related failures into incidents, then kicks off what it calls an investigation. The agent pulls in logs, traces, metrics, recent deployments, even old Slack threads if they're relevant. Once it has assembled enough context, it drafts a single pull request per incident and drops it directly into Slack. Engineers can merge it, dismiss it, or pop it open as a Claude Code session to edit the fix before shipping.
The system leans on vendor-neutral OpenTelemetry, which theoretically means you can keep your telemetry data if you decide to bail on Superlog later. It also tracks LLM costs by callsite, tenant, and model—useful if you're running AI-heavy workloads—and tags each incident summary with a confidence score. Whether those scores prove reliable in production is an open question.
The Team Behind It
Superlog's co-founders split responsibilities neatly. Arseniy Shishaev, the CTO, spent three years at Datadog building data tooling for historical metrics before starting a previous venture called Bluco. He's an École 42 Paris graduate and won national olympiad honors somewhere along the way. Nicolò Magnante, the CEO, comes out of Boston Consulting Group and claims to have scaled two startups to millions in annual recurring revenue, though those figures don't appear to be independently verified anywhere public.
They launched with support for Python, FastAPI, Flask, Next.js on Vercel, LiveKit agents, Expo/React Native, and Supabase Edge Functions. Generic coverage extends to Go, Java, Ruby, Rust, .NET, PHP, Elixir, and Node. Installation runs through an open-source skills repository—Apache 2.0 licensed—that lives on GitHub. Developers can run npx skills add superloglabs/skills --all and hand the project to an agent for onboarding.
The fact that Superlog is a two-person operation is worth noting. Building something this technically complex with that little headcount either signals extraordinary efficiency or suggests the product is earlier-stage than the polish implies. Perhaps both.
A Busy Month for AI Observability
Superlog's timing puts it shoulder-to-shoulder with several other AI-native observability launches. Honeycomb rolled out Agent Observability the same week, promising full visibility into agentic workflows. Middleware released OpsAI earlier in May, automating root-cause analysis and generating fixes. Selector announced AI-powered multi-cloud observability on May 19.
A market overview published on May 16 lumped Superlog into the "L2 agent trajectory" category alongside Honeycomb, LangSmith, Arize Phoenix, LangFuse, and Helicone. The report credited Superlog with the "lowest configuration burden" but noted its "very new, limited track record." That's diplomatic phrasing for: no one knows yet if this actually works at scale.
Superlog differentiates itself less on visibility—plenty of platforms already offer that—and more on closing the loop. The founders frame the product as "observability that's meant not to be opened." The idea is that if the system can fix things autonomously, engineers shouldn't need to spend hours in dashboards triangulating failures.
It's an appealing vision, especially for small teams drowning in alerts. Whether it holds up in messy production environments is another matter entirely.
The Trust Problem
The Show HN discussion laid bare the central tension. Developers asked for dry-run modes that would preview changes without touching live code. They wanted explicit inventories of which files Superlog modifies. They wanted clarity on telemetry egress before routing production traffic through the system.
These aren't idle concerns. Auto-generated pull requests are only as good as the context feeding them, and a poorly reasoned merge can cause more damage than the original bug. The stakes get higher when you consider that most codebases carry technical debt, undocumented decisions, and edge cases that even human engineers miss.
Pricing remains undisclosed as of late May. Customer details, too. The Y Combinator profile lists a team of two and no open positions. The website—superlog.sh—is live but wasn't accessible via standard web crawlers during research for this piece, an oddity for a product trying to gain traction. Interested teams can email [email protected] or poke around the open-source skills repository to see what the wizard does under the hood.
What Comes Next
The broader question isn't whether Superlog's specific implementation works—that will become clear soon enough as early adopters either rave or churn. The real question is whether the industry is ready to hand agents this level of autonomy.
Observability has always been about surfacing information so humans can make decisions. Superlog flips that dynamic. It wants to observe, decide, and act—leaving engineers to review the results rather than drive the process. That shift requires not just technical maturity but organizational trust, something startups with "very new, limited track records" don't automatically earn.
For now, Superlog represents a bet: that developers are tired enough of toil to let an agent touch their main branch, provided the fixes hold up more often than they fail. Whether that bet pays off will depend less on the elegance of the automation and more on how many times the system gets it right when it matters.
