iiinova — where your digital self keeps evolving

MartinDotNet @martindotnet.bsky.social · 06/10/2026

This pattern should be part of Aspire right?

MartinDotNet @martindotnet.bsky.social · 02/10/2026

What language? How can we help? This isn't the perception that a lot our customers have, some languages are harder than others. We have built some skills on some good practices though.

000

Reposted by MartinDotNet

AWS UG UK @awsuguk.bsky.social · 01/10/2026

The videos are live! 🎤 @theburningmonk.com - Event-driven architecture: the hard parts 🎤 @martindotnet.bsky.social - Every Player Gets a VM: Battleships on Lambda MicroVMs for re:Invent 👉 awsuguk.org/our-videos/ #AWS #AWSCommunity #Serverless #eventbridge #lambda

001

Reposted by MartinDotNet

Honeycomb @honeycomb.io · 14/09/2026

New O11yCast — Mike Goldsmith, Martin Thwaites, and Ken Rimple on adaptive sampling, including Honeycomb's new Tail Sampling Processor for the OpenTelemetry Collector. Have a listen here: honeycomb.io/resources/podcasts/ep-93-adaptive-sampling-strategies-with-mike-goldsmith

042

Reposted by MartinDotNet

AWS UG UK @awsuguk.bsky.social · 10/09/2026

65,000 attendees, four days, and a booth game that has to survive the moment a keynote ends. 🎤 @martindotnet.bsky.social, Every Player Gets a VM: Prepping Battleships for re:Invent Scale 👉 www.meetup.com/awsuguk/even... #AWS #AWSCommunity #Serverless #lambda

001

Reposted by MartinDotNet

AWS UG UK @awsuguk.bsky.social · 04/09/2026

AWS UG UK London 23 Sept 🎤 @theburningmonk.com - EDA: the hard parts: coupling solved, three new problems 🎤 @martindotnet.bsky.social - Every Player Gets a VM: surviving 65,000 attendees at re:Invent www.meetup.com/awsuguk/even... #AWS #AWSCommunity #Serverless #Observability #EventDriven #Lambda

001

MartinDotNet @martindotnet.bsky.social · 25/08/2026

The underrated part of "well-instrumented" is the "well-documented instrumentation". Document the attributes and spans in a way the agent can read them and it's goes for 2x to 10x useful pretty quickly. I'm doing this on my side project app, and it makes sonnet feel like fable.

031

Reposted by MartinDotNet

Ian Cooper @icooper.bsky.social · 11/08/2026

As developers shift left from coding to architecture, understanding how apps interoperate has never been more important. At NDC Oslo, I will be running my Practical Messaging Workshop, where you will learn the skills needed to design event-driven architectures at scale. ndcoslo.com/workshops/pr...

ndcoslo.com

Practical Messaging | NDC Oslo 2026

In this tutorial, we will look at distributed systems, and how we integrate them. We will understand why we would prefer to integrate via messaging, the fundamentals and key concepts of messaging and…

381

Reposted by MartinDotNet

Honeycomb @honeycomb.io · 04/08/2026

Log everything, structure nothing, then can't find what you didn't know to look for. That's the trap last week's Observability Masterclass focused on. @lizthegrey.com built wide events live with OpenTelemetry. On demand now. Next up Aug 19: feedback loops x observability. 🐝 go.hny.co/4wdFfDj

go.hny.co

O11y Masters: Observability Engineering Masterclass with Liz Fong-Jones

A six-session live masterclass with Liz Fong-Jones, turning Observability Engineering into practice. Aug 3 – Oct 14, 2026.

054

Reposted by MartinDotNet

Liz Fong-Jones (方禮真) @lizthegrey.com · 31/07/2026

Monday 10am Pacific / 1pm Eastern: the first session of my Observability Engineering Masterclass series goes live. 45 minutes, live demos & code. Recording plus a self-serve lab go out right after, so you can work through it again at your own pace or share with your team. New class every 2 weeks.

1315

MartinDotNet @martindotnet.bsky.social · 10/07/2026

If theres things missing that would allow better analysis in a backend, we will absolutely add them to the semconv, so that libraries can then pick them up. That could even be concepts involving baggage for advanced context propagation.

200

MartinDotNet @martindotnet.bsky.social · 10/07/2026

I see that slightly differently. The conversation is something that covers a user or a system having an interaction with an agent. Its still a conversation if both sides are a program. The attributes and data exist to do all the things you're talking about, its a rendering issue for backends.

100

MartinDotNet @martindotnet.bsky.social · 10/07/2026

Tying is an LLM call to its associated tool call chain is coming soon, its been proposed under "groups". I'm nit sure what else is missing? A trace covers the full call tree from a tool into the backend.

100

MartinDotNet @martindotnet.bsky.social · 10/07/2026

Can you give an example of what you mean here? Are you referring to looking at the thinking portions of reasoning models and capturing that separately? The idea is that a user interaction with an agent is a trace, the series of user interactions are a conversation.

100

MartinDotNet @martindotnet.bsky.social · 10/07/2026

We have those? Correlation across interactions is done through `gen_ai.conversation.id` and theres standards for all the attribute like token, messages, models etc.

200

MartinDotNet @martindotnet.bsky.social · 10/07/2026

Whats the missing bit of OpenTelemetry you don't have right now?

100

MartinDotNet @martindotnet.bsky.social · 03/07/2026

I was the same, then I figured "why not just join an Observability company and then its at least explicitly my job"

110

MartinDotNet @martindotnet.bsky.social · 02/07/2026

I do, but I use Claude and our MCP, not the product.

000

Reposted by MartinDotNet

Andrew Lisowski 💻 @hipstersmoothie.com · 21/06/2026

For @cocore.dev in the last day I - converted ALL of the backend code to idiomatic effect - added o11y and @honeycomb.io - squashed the top errors that were obvious Pre ai all that is a weeks long effort, some I’d even have to learn a lot (effect) Now it’s super high quality, no effort.

3201

MartinDotNet @martindotnet.bsky.social · 21/06/2026

You should checkout asking canvas to do the investigation then point your agent at canvas. Also keep an eye out for some of the stuff we'll be releasing next month, a load of cool canvas stuff coming up.

000

MartinDotNet @martindotnet.bsky.social · 21/06/2026

Did you use the MCP investigate the errors? Or Canvas Agent?

100

MartinDotNet @martindotnet.bsky.social · 15/06/2026

Thats just coding agents though, and to me they're the least interesting usecase for agents. Using agents as fuzzy automation of internal business processes is way more interesting.

100

MartinDotNet @martindotnet.bsky.social · 15/06/2026

The real tax that people forget is that every single prompt you send to the agent includes every previous prompt (and the thinking if you enable preserve_thinking), therefore reading a file to understand it will include that context even if you change the file.

200

MartinDotNet @martindotnet.bsky.social · 15/06/2026

In coding usecases, having an agent that understands LSPs, delegates summarisation to sub-agents with narrow context is the right answer. Claude Code isn’t actually good at that. Augment Code for paid, or OpenCode with tuning of the agents and prompts is better.

310

MartinDotNet @martindotnet.bsky.social · 15/06/2026

I think, right now, people are only thinking about coding agents running a conversation they're part of, and I'm finding way more usecases for agents right now, coding agents are the least interesting of those.

100

MartinDotNet @martindotnet.bsky.social · 15/06/2026

This is where getting effective at the prompts you give to the sub-agent is important. To be clear, I'm not talking about coding agents here. I'm talking about agents you build for people to interact with. Creating narrowly focused sub-agents is incredibly effective in that scenario.

310

MartinDotNet @martindotnet.bsky.social · 15/06/2026

Research is a good example of this. Pushing 2000 tokens into a sub agent for it to do web research over 20 turns, racking up 200k input tokens, only to pass back 10k tokens back to the main context that runs for another 20 turns, that saves a lot. Even more so if the subagent uses a lesser model.

100

MartinDotNet @martindotnet.bsky.social · 15/06/2026

That depends how you use them. If you want an agent that can investigate the state of an order through multiple tool calls with a defined end state, sub agents are the way to go. There are lots of use cases where the sub agent doesn't need a lot of context.

100

MartinDotNet @martindotnet.bsky.social · 14/06/2026

As always, the meta work is what people like doing. My theory is because there's less approvals, no product managers and EMs in the way. The same reasons as dashboards in general because they have full authority to build a UI that "people" use.

110

MartinDotNet @martindotnet.bsky.social · 14/06/2026

Working out whether sub agents will save you money on token usage and performance of agent responses. This the important thing. Slapping some otel environment variables on claude and building a dashboard, while cool, isn't anywhere near as useful.

120

MartinDotNet @martindotnet.bsky.social · 14/06/2026

Token usage in local coding, while important, isn't the most important thing in writing code. Token usage in your production app, being able to segment by user, user agent, intent (classification), usefulness (eval results), uncached vs cached. Thos are way more important.

330

MartinDotNet @martindotnet.bsky.social · 14/06/2026

I've seen people building elaborate visualisations around everything from token usage to prompt sizes. Not at an organisation level for cost management but for their local claude coding sessions. However, when you suggest they add some attributes to their production app, apparently you have 2 heads

130

MartinDotNet @martindotnet.bsky.social · 14/06/2026

I find it hilarious that engineers will spend days, weeks even, curating their token dashboard for their claude code session, when they'll spend practically zero time adding telemetry to their apps to make it better. Claude Code: heres 72 graphs and dashboards. Production: install agent, go home.

390

MartinDotNet @martindotnet.bsky.social · 14/06/2026

Try doing that with the new Canvas! Then you'll get a full visual representation of everything. Then ask your coding agent to ask the canvas for the conclusion and ask it to implement the fix. This way you end up with a persistent sharable artifact for the investigation too

020

MartinDotNet @martindotnet.bsky.social · 11/06/2026

I will happily donate my time to jump on a call about this stuff, I can even bring one for the JS otel maintainers with me. Theres a lot of words there that don't make sense to me. I only know the pains of JS otel right now and would love you not to get that hate for something awesome.

020

MartinDotNet @martindotnet.bsky.social · 11/06/2026

Have you done any validation to ensure that you're emitting enough information to be otel compliant?

100

MartinDotNet @martindotnet.bsky.social · 11/06/2026

"Someone" will need to build that otel receiver, which we're increasingly relying on authors of frameworks to build those out as they're the ones that know where attributes come from etc. The hope is that if the author is the one doing it, then the telemetry available is standards compliant.

100

MartinDotNet @martindotnet.bsky.social · 11/06/2026

I would assume that nuxt is producing a first part otel integration though? Based on this API. Thats the real value these days, not APM tools.

100

MartinDotNet @martindotnet.bsky.social · 11/06/2026

I've assumed OpenTelemetry here, not APM tools. Maybe that isn't correct?

110

MartinDotNet @martindotnet.bsky.social · 08/06/2026

Not always new telemetry, most of the time the telemetry exists, you just didn't know it was there, or couldn't analyse it in the way needed. Otherwise yes, that is how people have worked for a few decades. "We can't get how many users hit this endpoint" "We need a metrics" But the data exists

010

MartinDotNet @martindotnet.bsky.social · 08/06/2026

Thats still building things with the assumption that you know both the question that needs to be asked, and also how that question needs to be answered. That requires a certain level of clairvoyance around how the system evolves over time, and how production usage scales.

110

MartinDotNet @martindotnet.bsky.social · 08/06/2026

The key is "Correctly Attributed" Blame ;)

110

MartinDotNet @martindotnet.bsky.social · 08/06/2026

Oh boy... Have I got the demo for you :)

010

MartinDotNet @martindotnet.bsky.social · 08/06/2026

Yup, in 2026, dashboards are become less relevant. The issue we saw with MCP centric analysis is that visual correlation and investigation is still important. The AI missing things that a human might catch during visual analysis. Its about how we surface that better, in an investigative way.

110

MartinDotNet @martindotnet.bsky.social · 08/06/2026

The problem is thinking that dashboards are the goal, or even useful. The goal is understanding, to get there you don't need static things, you need dynamic exploration. Dashboards show you what you understand about your system already, they don't generate new understsnding.

110

MartinDotNet @martindotnet.bsky.social · 08/06/2026

Because its "just a database and some stored queries". Forgetting that observability isn't about the data, or the dashboards. When you look at observability as dashboards, metrics and log search, it really does sound pretty easy and overpriced.

110

MartinDotNet @martindotnet.bsky.social · 08/06/2026

Viewing it as a way to get shared ownership is a good way to think about it. Blame is the negative said, but ultimately its about getting "past" the blame stage quicker. Its human nature to say "who did this", if only to go to the source for more information. If thats "everyone" its quicker.

120

MartinDotNet @martindotnet.bsky.social · 03/06/2026

I have not, at least I don't think i have, remind me next week, happy to have look

010

MartinDotNet @martindotnet.bsky.social · 01/06/2026

Prompt processing is what I'm more interested in.

100