Since then, I’ve been pretty deep in the LLM/agent dev tools, and observability has been my main thing.
I ran a small poll on LinkedIn recently about where teams are with observability for LLM-powered apps/agents. Results:
• 20% instrument LLM observability from day 1 • 30% plan to implement later • 20% are building an in-house solution • 30% are still learning about this space
That 20% building in-house was the most interesting to me, so I followed up with a mix of early-stage, YC founders and more mature orgs. The drivers I kept hearing:
1) Local / self-hosted models Some teams assume there aren’t viable observability options for local/hybrid LLM stacks, so DIY feels like the default. In practice, there are ways to do this, but they’re easy to miss right now.
2) Cost uncertainty Token usage is hard to estimate early on, so pricing feels unpredictable. A minimal in-house layer looks safer than surprise bills.
3) Control + speed Bootstrapping basic tracing/logging is straightforward and gives full ownership while teams iterate quickly on the core product.
This reminds me a lot of early APM / product analytics. Many teams started with “we’ll just implement our own logging.” Totally reasonable at the beginning — but once usage and complexity scaled, that logging quietly turned into:
• an internal platform to maintain • a backlog of features to build • a growing surface area of edge cases to debug
…often becoming a real distraction from the core business.
Our bet is LLM/agent observability follows the same path: teams start with DIY logging, then realize it’s becoming a side-product, and eventually most adopt a standard platform early. We’re also seeing APM/analytics vendors expand into LLM flows, which reinforces that direction.
What we’re building My team and I are working on LLM/agent observability focused on usage, cost/pricing, evaluations, and debugging. Most teams we talk to still don’t have anything in place, even when LLMs are core to the product, so we’re trying to make the “day 1” setup practical.
We're part of a larger org, but this team is being run like a startup within it: small group, fast cycles, heavy on user conversations, and shipping quickly based on real usage. That setup is why we’re doing early access and iterating closely with teams.
Early preview / notes here: https://aback-handbell-1cd.notion.site/Progress-Observabilit...
We’re planning to support self-hosted options as well.
If this is relevant to what you’re building and you want to help us shape the LLM Observability you need, we have a free Early Access Program here: https://www.telerik.com/agent-observability-early-access