Subtrace Wants to Show You What Your AI Agents Are Actually Doing

The Macro: Your AI Agents Are Black Boxes in Production

Here is the situation most AI teams are in right now: they built an agent, it worked in testing, they shipped it, and now it breaks in production in ways nobody can explain. The agent called the wrong tool. It hallucinated a function parameter. It looped three times on the same API call and burned through the rate limit. And the logs? The logs show you that something happened. They do not show you why.

Traditional observability tools were built for a world where services make predictable HTTP calls in predictable patterns. Datadog, New Relic, Grafana, they all assume your application behaves roughly the same way every time a request comes in. AI agents do not work like that. An agent might take four different paths through your tool stack depending on how it interprets the user’s input. The call graph is nondeterministic. The failure modes are nondeterministic. And the existing monitoring stack was never designed for this.

The market for AI observability is still young. LangSmith handles tracing if you are locked into the LangChain framework. Arize Phoenix does model monitoring with a focus on embeddings and drift. Helicone tracks LLM API usage and costs. But there is a gap at the infrastructure level, the layer where HTTP requests actually happen, where tool calls resolve, where you can see the raw behavior of an agent without framework-specific instrumentation.

That gap is where Subtrace is building.

The Micro: A Zero-Code Network Tracer Built by Google and Microsoft Alumni

Subtrace is a network-level tracer that captures every HTTP request your AI agent makes in production, without requiring any code changes or root privileges. It uses BPF (Berkeley Packet Filter) technology to hook into network calls at the kernel level, which means you get full visibility into what your agent is doing without instrumenting a single line of application code.

The pitch is straightforward: deploy Subtrace alongside your agent, and you get a real-time trace of every tool call, every API request, every external service interaction. When your agent fails in production, you can replay the entire request chain and see exactly where things went wrong. No SDK integration. No framework dependency. No code modifications.

Adhityaa Chandrasekar and Sachin Sridhar founded the company in 2024. Adhityaa previously worked on Kubernetes internals and GKE infrastructure monitoring at Google. Sachin built data infrastructure at Microsoft, working on Power BI and Microsoft Fabric. Both are IIT Madras graduates. The technical backgrounds are directly relevant here. BPF-based tracing is systems programming, and having people who worked on Kubernetes observability and large-scale data infrastructure is exactly the kind of experience you want behind a product like this.

They are a two-person team in San Francisco, part of YC’s Winter 2025 batch. The project has picked up roughly 2,500 GitHub stars, which for a developer infrastructure tool at this stage is a strong signal of organic interest.

The product sits at an interesting intersection. It is framework-agnostic, so it works whether you built your agent with LangChain, CrewAI, AutoGen, or your own custom stack. It runs at the network level, so it captures behavior that application-level tracing misses. And it requires zero code changes, which removes the biggest friction point in getting observability adopted by engineering teams who are already stretched thin.

The Verdict

I think Subtrace is solving a real and growing problem. Every company shipping AI agents to production is going to need observability tooling that actually works for nondeterministic, multi-step workflows. The current options are either framework-locked or focused on model performance rather than runtime behavior. A zero-code, network-level tracer fills a genuine gap.

The challenge is timing and positioning. The AI agent infrastructure market is moving fast, and the big observability platforms are not going to ignore this space forever. Datadog already has LLM Observability in preview. When the incumbents show up, Subtrace needs to have built enough depth in agent-specific tracing that switching costs matter.

Thirty days from now, I want to see how many production deployments they have and whether teams are using Subtrace for debugging or just for monitoring. Those are different use cases with different retention profiles. Sixty days, the question is whether they have expanded beyond individual developers to team-level adoption. Ninety days, I want to know if they have found a repeatable sales motion or if every customer is still a hand-held onboarding. The technology is sound. The founders have the right background. The question is whether they can build a business around it before the window closes.