Aedifico

I used to talk to one agent who orchestrates other agents, and it costed still alot of time to build stuff. I was definitely faster, but I was too much the bottleneck, and I had to manage alot of things by myself, because instructions are not always followed.
So I decided to build AgentFlow, my companion in building things on the go. It uses static workflows to streamline development, and enforces the agents to cheat less.
So now my work goes through fixed pipelines with agentic steps, where each agent has 1 shorter goal, and orchestrated automatically, and verified ruthlessly. So which agent is executed, must be executed stuff are done by the pipeline, so agents can be creative, and fill in the stuff in their boundaries.

2026-07-29 08:21:14 UTC

- reply

Whats up with this?

2026-07-29 06:14:19 UTC

- reply

In what sense?

2026-07-28 19:11:42 UTC

- reply

Would be a better marketing for it 🍻😂

2026-07-28 13:53:08 UTC

- reply

Thats a good sign! 🤜🤛 I am gonna run today after not running for a while

2026-07-28 12:38:16 UTC

- reply

Anything fun care to share?

2026-07-28 12:37:39 UTC

- reply

Lets go! 🍻

2026-07-28 12:37:14 UTC

Are you working on your top 3 pain points?

2026-07-28 12:36:19 UTC

"I rebuilt an app in an evening" 😂

2026-07-28 10:52:13 UTC

- reply

Yeah, but I mean, is it better to ship 4x more, if the price is to be glued to your machine? Is it the right type of productivity?

2026-07-28 10:50:20 UTC

Is it productivity when you are glued to your machine?

2026-07-28 10:03:26 UTC

I did some research on industry best practices for agentic coding/agentic work, skill repos, tools (tenex, what I could find from wingman, etc), so I can create my ideal workflow on the go.
I found that hard gates, and automatic execution of verification/steps (not simply giving it to the LLM to do it, or encoding it in skills), is not used that much in many frameworks.

2026-07-28 09:57:40 UTC

Instructed to do and does are two different thing.
great good morning

2026-07-27 11:07:31 UTC

- reply

Arent they like words? Words are basically nothing either, but we get offended so easily.

2026-07-27 10:55:43 UTC

- reply

I would argue the best things in life are not even for sale.

2026-07-27 10:54:22 UTC

It is a really great feeling when you put effort into something, and it is blooming.
great good morning

2026-07-27 10:47:01 UTC

- reply

On the roadmap. Also benchmarking TEE models for my app.

2026-07-27 07:42:08 UTC

- reply

As per an article/comparison I read, its cost is close to GPT 5.6 Sol, and half the price of Fable 5 based on multiple tasks.

2026-07-27 06:23:57 UTC

- reply

The aunt workout

2026-07-25 13:11:06 UTC

- reply

How do you manage the orchestration?

2026-07-25 09:41:44 UTC

Simply using harnesses (claude code/codex/opencode/etc) or using a workflow engine?
I vote for workflow engine. What must be done has to be executed and not trust the mercy of an LLM.

2026-07-25 08:56:21 UTC

Put the bar so high, that they will remember you.
As a first step, you can just pay attention more.

2026-07-25 08:54:51 UTC

- reply

How did you make sure it did not share secrets from your server, and the end result is more secure?

2026-07-25 08:53:16 UTC

- reply

Great day! PV! ☕

2026-07-24 20:07:06 UTC

- reply

How do you make sure that the AI does not cheat, like hacking huggingface? Would be curious about your setup. I am building now a really simple workflow with just. I took the time today to design my dream setup.

2026-07-24 19:35:48 UTC

- reply

buzz is interesting, with the collaboration, but does not seem to follow the industry best practices for output verification. I guess it could be good for non-verifyable type of work, like marketing, design, brainstorming. But where a workflow that does automated verification is better, and automated execution is important, it lacks. But cool to see it anyway! I would only advise taking it granted as the solution for all type of ai collaboration/team work.

2026-07-24 14:35:26 UTC

Any experiences with spec-kit?
#qa #speckit

2026-07-24 10:50:57 UTC

- reply

Enjoy!

2026-07-24 10:43:53 UTC

This could be a nice joke in a movie, but...
#nevent1q…z8kh

2026-07-24 07:02:19 UTC

- reply

🤜🤛

2026-07-24 06:57:31 UTC

- reply

Also how can you handle multiple session with it? Not to bankcrupt your credits for huge context?

2026-07-24 06:54:37 UTC

- reply

Not sure you have it, or works for you, but add an in-app update into your apps, so you can download the new versions of your apps on the go, and test it. Its a game changer for me. Talking to it, sounds the next.

2026-07-24 06:51:19 UTC

- reply

It reads back the answers as well? So fully handsoff?

2026-07-23 18:17:35 UTC

wamp server 🫠

2026-07-23 15:44:05 UTC

What is your moat?

2026-07-23 15:38:17 UTC

- reply

playwright cli

2026-07-23 15:34:35 UTC

Its interesting to ponder about your unrealized successes.

2026-07-23 15:31:30 UTC

Its good to close a chapter in your life. It reveals the missed opportunities. You can learn, and move on.
If you had stayed in that chapter, you would probably not be aware of the missed opportunities.

2026-07-23 15:28:26 UTC

- reply

A lot of small steps are actually big steps.

2026-07-23 15:25:17 UTC

- reply

What is wisdom anyway if not distilled experiences?

2026-07-23 14:19:28 UTC

- reply

money != time
But it gives you time. Ironic.

2026-07-23 10:44:51 UTC

Start doing. You have more tools now than ever to learn by doing.

2026-07-23 10:01:16 UTC

"Markets will wear you out or scare you out"
I think this is more broadly applicable for real life.

2026-07-23 09:59:19 UTC

What are you using local models for on your phone? And which models?

2026-07-23 05:42:30 UTC

#HealthOS is a bit heavy on model size.
https://blossom.primal.net/34bd59b19b783a45a1cc5b29e23767c21d56984bee3834305cc21ac5803b4674.jpg

2026-07-23 05:30:46 UTC

Tool calling with local gemma 4 model is an interesting topic. By default it is a non-tool call supporting model, but it is possible to make it tool calling, if we use it to generate text, and parse the tools from that. Thats what Google does in Edge Galery.
I am testing now different models with using the features of #HealthOS so far it seems QA functionality with tool calls work with 4B models.
great good morning.

2026-07-22 20:19:14 UTC

https://youtubesummary.com/

2026-07-22 20:18:39 UTC

#HeathOS has a decent offline voice to text now after industry best practices. Not an easy job though on a phone hardware without cloud models. I also added a private PPQ cloud cleanup, which costs like 60 cleanup "jobs" for 1 cent, or so.
Next is evaluation of LLMs for agentic usacases to find the cheapest still performing model. I guess coaching will still need bigger models. Lets see, let the eval decide.

2026-07-22 19:35:03 UTC

AI is a tool. Like photoshop, you can do awesome things with it, but you need to know how to use it properly.

2026-07-22 19:34:07 UTC

- reply

Exactly. 😂

2026-07-22 19:33:02 UTC

- reply

How about having a "coach"?

2026-07-22 19:32:18 UTC

Developing an app is simple. But not easy.
Quality work is still effort. Just less than before.

2026-07-22 18:45:53 UTC

They are similarly capable, cheaper, but could run in a TEE. No brainer.
But subcriptions are still a thing if you want to get more bang for your buck. But subscription costs seems to be hard to maintain for long, and then proprietary models will have a price disadvantage too. Not mentioning the privacy argument.
#nevent1q…sqw2

2026-07-22 05:31:09 UTC

It turns out that claude code has a "turn veto" feature, that does not allow the model to finish if it does not pass the tests.
And it is not implementable as a real veto in opencode or pi harnesses. You can approximate it, but it is not a real veto.
A good harness with good hooks is really important for quality work.
What hooks do you use?

2026-07-22 05:25:42 UTC

The process is part of the product.

2026-07-21 11:39:45 UTC

The all-in-one app, but privacy aware.
Why this isnt a thing?

2026-07-21 10:37:22 UTC

#HealthOS
Your private health intelligence. Local first. Encrypted. No compromise!

2026-07-20 14:45:32 UTC

Running private mesh with headscale+tailscale+gluetun+mullvad
It is incredible how complex, and not a supported feature it is to
1. Reach your own devices
2. Use only a wireguard egress to access the internet.
3. Change egress/exit ip from your phone.
4. Works with android block all non-vpn connections.
5. Works on windows next to a vpn app.
So many nuances, and there is no built solution yet. I feel like, noone wants this?

2026-07-20 14:37:30 UTC

- reply

I guess the controversial stuff is rather image contents are using servers too.
This is how the whole pixel tracking work.

2026-07-19 15:58:10 UTC

Build many apps. Thats the new way of learning.
In the past, we had to selectively build stuff, because it took a lot of time. Now we can just build more, and learn from its mistakes.

2026-07-19 11:21:11 UTC

bulding with private/proprietary models == giving unfair advantage/knowledge to private companies to build your idea
Probably not a sustainable way to make new businesses.

2026-07-19 11:15:53 UTC

Kimi K3 seems to flip the frontiers-are-only-private-models

2026-07-19 11:05:53 UTC

- reply

Either that, or banned, or your account locked.

2026-07-19 11:04:46 UTC

- reply

Do you have task benchmark results? Or this is from using it on probably comparable tasks?

2026-07-19 06:15:02 UTC

Knowing, believing and acting on something are 3 different things.
great good morning

2026-07-17 19:34:59 UTC

- reply

Because of the human weakness to fall for "easy" wins. People want to win more with less.

2026-07-17 15:08:40 UTC

Fun building a local first app with a dumb encrypted relay, and bricking the local database 🤣🤣🤣

2026-07-17 10:22:37 UTC

- reply

big leap

2026-07-16 15:28:48 UTC

I am building a "private brain for your health".

2026-07-16 14:17:09 UTC

- reply

If you are talking about in-transit emcrypted data, sure. What I meant is, if each phone becomes a FIPS node, and they have internet, bluetooth, wifi to connect to other nodes, it has a bigger reach of devices, so possibly bigger damage is possible. But not inherent from the design.

2026-07-16 13:46:33 UTC

- reply

Would be nice and dangerous too 🧐 If its a fips node, viruses can also propagate better through the different transfer mechanisms. Like the zero day series. But I dont say it is not already possible with apps million people use though.

2026-07-16 12:43:21 UTC

- reply

I love the versatility of your attention Kat! 🫂

2026-07-16 12:41:47 UTC

Actually it is quite amazing that you can be an app dev/android dev/ios dev if you have the persistence to it, just in a few days, and you can learn so much.
Learning by doing is the new norm. Lightning fast.

2026-07-16 12:37:24 UTC

- reply

I would vote for FIPS first, and then sell servers and compute over FIPS.

2026-07-15 16:56:48 UTC

- reply

auto allow?

2026-07-15 16:53:47 UTC

Health apps shall be local first.
You cannot change your health from one day to another, and it influences your life quite a lot, so it shall be shared with the least amount of services.

2026-07-14 19:39:33 UTC

You are only an expert of a topic, if you can explain it to 5 year olds.
And if you are such an expert, take it as a job, and help people in that area.

2026-07-14 15:52:42 UTC

- reply

Unless the differentiating factor cannot be copied.
gm

2026-07-14 09:21:30 UTC

- reply

Hmm 🧐☕
https://blossom.primal.net/e9d387619cee6ac63b139fc03b11f816f8e923fb20e123aad0982101862feb57.jpg

2026-07-14 08:15:50 UTC

- reply

I love this thread. LLMs realized it sooner than most people.

2026-07-14 08:10:22 UTC

Best app marketing when real discussions casually lead to showing something in your app.
Without you saying: "I am developing this app blabla...".

2026-07-14 07:50:18 UTC

When you get the hang off building for your own need, eveything changes.
great good morning ☕

2026-07-11 14:27:57 UTC

- reply

Whats your "proper remote setup"?

2026-07-10 05:41:52 UTC

good morning,
rise and shine

2026-07-10 05:40:04 UTC

- reply

Experiences so far?

2026-07-09 19:28:43 UTC

Implementations, where data has to be shared with a third-party/service provider, when it could be done in a smarter or more local way, are so annoying.
Lets change that!

2026-07-09 19:27:15 UTC

I want to live in a world, where applications are private by default, does the processing on device, cloud processing are optional, secured and verifyable, and provides high value in return.
That is why I am working on "#HealthOS the allrounder health companion and coach app", that helps in every part of health, and it is local by default.
My passion is about creating a healthapp, that gives its users the control. Not ostensible, but real control.

2026-07-09 13:18:10 UTC

- reply

I guess you can also burn a ton of money being an ai provider

2026-07-09 08:23:15 UTC

- reply

Interesting take! What would you attribute this to? What advantage that allows him to overtake others?

2026-07-08 19:58:55 UTC

All it takes is to stay calm, and concentrate on solving problems instead of raging.

2026-07-08 19:11:56 UTC

Who invented scented trash bags? 😂

2026-07-08 15:29:35 UTC

Plus added on-device voice-to-text, and text cleanup with AI.
So now I can use private models for text cleanup.
#nevent1q…36qm

2026-07-08 15:28:26 UTC

#HealthOS supports ppq TEE models.

2026-07-08 14:22:44 UTC

- reply

I love how well relaxed some people can be. Why so much stress! 😅🤣

2026-07-08 14:21:58 UTC

- reply

Every action has a reaction 🤣

2026-07-08 10:21:07 UTC

#HealthOS in-app agent at the moment can generate a personalized mealplan for $0.09 with gemini 2.5 flash.
Would you pay 9cents for a mealplan tailored for your nutrition needs and recipes?

2026-07-08 10:17:12 UTC

I am fine-tuning my #HealthOS in-app agent for personal meal planning.
gemini 2.5 flash works good. It is cheap for testing, and I can improve its result by fine-tuning the agent harness files.
Next will be to convert these manual tests into an automated eval, so I can compare multiple models with a benchmark cost vs result quality.