Thank you! super happy that's how you feel about Superlog. Let us know if you want to try it out and/or have any feedback :)

Great question! The investigation agent has access to all the telemetry - not only one service. So we can actually trace the root cause in such complex cases!

There are good ways to link operations between different services with OpenTelemetry (for example, passing the parent trace id in an inter-service HTTP/gRPC request). It's a bit tedious to do by hand, that's why we're publishing the skill that does that for you.

And totally agreed on config changes and deploy info. We've seen that having good environment and version control (commit hash, file name, line number) tagging is extremely important for root cause analysis, so we go hard on this in the skills.

We also have many infra integrations in our roadmap to make sure that we can deeply analyze the infra/config side of things.

Thanks for the feedback!

When you're installing Superlog, you can use any coding agent you'd like, including a local model.

Your telemetry then goes into our data stores, and right now we have one DC on the US west coast.

Whenever there's an error log or trace, Superlog can analyze it and prepare a resolution PR (or a note if something needs to be done manually).

This can be turned off and then the incident can be sent to your own models via a webhook.

We use one of the frontier models for that (it's an upstream AI provider). We're working on our own fine-tuned version of a SoTA model to minimize dependency on other AI providers.

To investigate an incident, we clone the repo in our worker, and pass the repository files to a coding agent in a sandbox. The agent has an MCP that gives it access to the telemetry (logs/metrics/traces) of the project.

The coding agent will then investigate the incident and prepare a patch. It hands over the patch via a tool. The worker then deterministically pushes the patch to a branch and opens the PR.

This way the agent doesn't have full Git access and can't do anything it's not supposed to do in the repository.

Awesome, let's definitely have a chat! I'll shoot an email via BF :)

Thank you! Glad you liked the install process :)

I'm afraid a collector and the operator are still the recommended way to go by OpenTelemetry (https://opentelemetry.io/docs/platforms/kubernetes/getting-s...). We're still working on a custom skill for Kubernetes, but the general skill should give you a sane default already.

A good way to start can be to start sending traces/logs directly by instrumenting the service and putting our backend as the collector.

I also help out personally whenever our clients have any questions on setting up the telemetry :)

Very good point on the confidence gate! We've rolled out feedback collection features on PRs themselves, on Slack notifications and incidents a few days ago so the data is still a bit fresh.

Anecdotally, our top clients accept 80-90% of PRs, with several clients accepting all of them and requesting an auto-merge feature. I myself accept most of Superlog PRs to Superlog. Most PRs that stay unmerged are usually due to a client losing interest in our product / abandoning the instrumented project.

Another interesting point is that not every defect is a PR. Often it's misconfiguration in an external service, so there's a special incident state for that. For example, yesterday I forgot to verify our domain on Resend so some verification emails didn't go through. Superlog pinged me on Slack and explained where to go to fix.

Super glad you like the npx onboarding and the MCP tool :) Please keep the feedback coming!

Thank you! Very good point.

Right now, the prompt will enumerate all the services and install the OpenTelemetry SDK (https://opentelemetry.io/) in each service.

Then for every service, the skill will make sure that:

- Every time something breaks and an operator needs to take a look, there's an error log - All important steps in a process emit info/debug logs (so that an issue can be investigated) - Operations are covered with spans with relevant attributes. - Cost (LLM tokens), API performance (latency/RED), tenant activity (cost/usage per tenant) are covered by metrics so that you can use Superlog MCP to build cool dashboards.

For most common stacks like NextJS, FastAPI, React Native/Expo etc. we have a custom skill that explains the best practices for this specific technology. For all the other stacks we ask the agent to use general best practices.

We have evals for all custom skills where we start from a starter project, run the agent with the skill and use LLM-as-a-judge to compare it to a human-written 'golden patch'.

In general, we try to:

- minimize diff, so that the instrumentation is easy to review - make small chunks of additive diffs vs huge indents / moving logic around - minimize new dependencies - use well-supported and audited OTel SDKs vs custom libs

You can read the skills here: https://github.com/superloglabs/skills.

I'll make sure to add this to our landing and print this out as the agent writes the code!

Thank you for the feedback!

Hm, sorry about that!

I made the Slack onboarding step mandatory for now since we thought that a lot of our value was in sending investigations and PRs, and Slack was what we used ourselves.

What tool do you use for communication around your project? If you don't want to share publicly, could you please shoot a line to:

ash [at] superlog.sh?

Would love to learn about your usecase in more detail too!

Very good point, thank you! Let me remove this phrase, you're right, it's misleading.

The pricing is only by usage (traces/logs/metrics) and investigation credits. We don't charge extra for repos :)

Good question! We don't have one as of today, just because we're iterating very quickly and a cloud version is the quickest way for us to keep things lean and up-to-date, but we're not far from having one.

Could you please send me an email at ash [at] superlog.sh? I'd love to hear more about your use case - we might have something for you very soon!

You're right! It's a big issue and I don't think there's a silver bullet.

We have an eval suite with code+telemetry fixtures and a golden RCA+patches and an LLM-as-a-Judge. So whenever we get feedback from our users and they're OK with it, we use their feedback to create an eval case (it's still quite manual since you have to calibrate the case).

We use Superlog to observe Superlog, so I often extract cases from our own errors. The PRs get better and better, but, of course, it's sort of a continuous improvement process.

Thank you! :)

Great question! I like to think about this in two ways:

1. Counter-positioning. Most existing tools have invested heavily in their web platforms and compete on their UI/UX. But actually, what matters to our clients is that bugs are fixed. Our top clients would rather never open our tool at all. If our competitors want to beat us, they essentially have to fight against their established business models that hinge on users looking at their browsers.

2. Evals. In order to have the most accurate RCA analysis you need a very good suite of evals: what was the right root cause in this bug? what is the right fix?. We're investing into this heavily, and as one of the early movers we have a big advantage here.

At the same time, I tend to approach strategy with a lot of caution. A lot of the canonical reasoning behind 'startup positioning' is based on extrapolation from trends, but surprisingly few analogies work in economics.

Our focus right now is: - talking to our users - making sure they have the best experience

Love the analogy! We honestly just wanted to have this product ourselves, and that was our primary motivation behind building it.

I agree with the philosophical principle! If you give a rigid observer an incentive to 'remove bugs', it will happily silence all alerts and report success.

Our goal is to make sure that doesn't happen. The investigation agent is actually a separate agent with a separate goal.

In practice, we rarely see the agent just silencing stuff. When this happens, I get on it and make it an eval case :)

How does a grep or read affect the observing system?

I guess the change in voltages, arrangement of registers, filling of buffers in the network stack are changing but... what?