TraceValut tracks every interaction and every decision the AI Agent makes in your codebase, storing it centrally.
@kgrajek.bsky.social talks about how we're making that data useful:
virtuslab.com/blog/ai/inte...
Safe agentic code gen: the goal of Safe Scala, presented by Martin Odersky during @scalar-conf.com
Waiting for the videos, here's a quick intro into what Safe Scala is, what are the alternatives (if any), and how to get started.
Is this the future of AI agents?
virtuslab.com/blog/scala/s...
Follow @rustikon.dev and @scalar-conf.com not to miss the next editions!
Two conferences, same question: how do you write code? Fully manually, everything is AI-generated, or somewhere in-between?
#Rust people seem more orderly, but also quite AI-sceptic. #Scala people on the other hand - a bit chaotic, but embracing the change.
See how Sandcat: Docker & dev container setup for securely running AI agents in `--dangerous` mode, works in a live demo!
In Sandcat, all container traffic is routed through a transparent mitmproxy, enforcing network access rules and injecting secrets.
www.youtube.com/watch?v=83Wm...
For me the biggest take-away is how AIs are different from juniors or onboarding a new person to a project.
A fresh developer knows, that they don't know. An AI doesn't know that it doesn't know. That's why it can be so much more confident in making the wrong change.
Great article by @askowronski.bsky.social on the different levels of context that you can provide to an AI agent. Which levels do you have at $work?
virtuslab.com/blog/ai/the-...
Can you cut your cloud bill by 50% without sacrificing performance?
One of our clients, a global leader in cybersecurity, did exactly that by migrating their massive EDR system from GCP to Oracle Cloud Infrastructure (OCI).
See how we helped them with this high-stakes migration:
Through a series of experiments, we've prepared a direct-style #Scala Skill for Claude Code.
We wanted to make sure that it only contains what's really necessary, and to help Claude where it stumbles. Even with a 1M window, context remains precious!
ElasticMQ - an in-memory @awscloud.bsky.social #SQS implementation for testing - just got a new UI!
Thanks to Michał Ossowski for his continued stewardship of the project and new development :)
Our client engagements often start with workshops on architecting the optimal setup for #Kafka clusters.
@mmatloka.bsky.social wrote down some of the basics of the process - still a long read, but who said disaster recovery is easy?
softwaremill.com/guide-to-apa...
Of course, unrecoverable errors at one level might become recoverable at a higher one - by catching them. But the only information then would be what's accessible at the point where the error is caught.
If there's any control flow you want to do - use a typed error. Otherwise - you get an untyped error-blob, which you can log, and respond to the user with whatever format of internal error that is appropriate. But there are no actions dependent on the error type.
Unrecoverable errors would be untyped **on purpose** (unchecked exceptions / panics) - making implementing control flow basing on the specific error as hard as possible.
* unrecoverable errors. This includes bugs, but also anything that should terminate the current "processing unit": respond with a 500 to an HTTP request, fail an incoming MQ message, or just terminate a CLI app.
Problems with error handling and exceptions, part N+1; is that a sensible way to categorise and implement error handling?
* recoverable, "expected" errors. Fully typed, as an `Either[E, T]` or `Result<T, E>` or a checked exception
Proteus - Tapir-style #gRPC endpoints in #Scala, by Pierre Ricadat, just got #Ox integration for streaming responses! All in direct-style, leveraging #JVM's virtual threads.
Check it out :)
github.com/ghostdogpr/p...
Just got my copy of DDIA by @martin.kleppmann.com, 2nd edition - one of the best books if your interests lie anywhere near distributed systems.
Proud to see @softwaremill.com among the acknowledgments! Even if our contribution was only financial, it's great to be able to support such works.
[1] github.com/softwaremill...
[2]
tapir.softwaremill.com/en/latest/se...
[3] sttp.softwaremill.com/en/latest/ba...
[4] github.com/VirtusLab/sa...
Some of the updates shipped during or right after:
* an example of direct-style response & request body streaming in Tapir [1]
* direct-style AWS Lambda handler [2]
* sttp+ZIO+Scala Native backend [3]
* hardening & naming consistency in Sandcat [4]
And more coming up ...
Best way to get a feature implemented or bug fixed? Talk to us @scalar-conf.com ;). Plus, our contributors are as active as ever!
We're closing the #Scala survey 2026 tomorrow, 26th March 2026, 23:59 CET. If you haven't yet, now's the time to take part :)
virtuslab.typeform.com/ScalaSurvey2...
Full paper here: arxiv.org/pdf/2603.00991
Another approach to consumer-agents-at-scale. Instead of sandboxing, verify that the agent stays within limits. LLM generates code on the fly, a compiler verifies it.
That's the approach described in the recent "Tracking Capabilities for Safer Agents" paper by Martin Odersky et al
Learning programming seems to hold its course so far. It's a good base skill, but how to evolve education with AI coding? Or maybe it's something that should be introduced only later?
We're steadily getting waves of registration on Shelly - our educational programming language for drawing. (I can only guess that each "wave" is a student group doing assignments.)
