spent the weekend so far documenting my life while exploring what web-native storytelling might look like for a 33 year long story.

im kinda satisfied with it.

if you're up for some weekend reading,

tej.as/story

i hate the internet sometimes

i feel very left out when i dont get early access to things like fable 5.2

man i have built something so damn powerful

daiku.ai

goated city tbf

I'm teaching a live remote workshop: Build a Production Agent Harness.

Mon 23 and Thu 26 Nov, 16:00 to 19:00 CET. Tools, context, evals, guardrails and observability, hands-on.

15% off with my link: ti.to/gitnation/ai-c…

It’s WLAN in German

What are you trying to fix in your agent stack?

Build a harness: @TejasKumar_ / @IBM
Review and repair PRs: Elise Gonzales + @bradfordc / @databricks
Build, verify, ship: @kentcdodds

AIE CODE workshops. Nov 10, SF.
Workshop add-on required.
ai.engineer/code/schedule

wow codex cowork is basically agi

We're giving away an iPhone Duo. 📱

Plus $200 Amazon gift cards and up to 80% OFF coupons up for grabs this Black Friday.

To enter:
✅Follow @Tenorshare_TS
❤️Like + 🔁 Repost
🎰Spin the Lucky Draw→ tenorshare.com/sales-promotio…
Ends Dec 1, 2026. Good luck 🍀

#Tenorshare #iPhoneDuo

i wrote a deep dive on software factories in practice and how i operate mine.

if ur uncertain at all about this term, this post is for u

👉 tej.as/blog/source-si…

everyone wants to move fast until it’s their turn to move

BTW, I did all this using a new tool from @TejasKumar_ called Daiku. The one sentence pitch is: "It gives your bots a browser." Of course, this unlocks everything since the world is still so web-based.

It also did my most recent airline check in.

Try it: daiku.ai/?ref=bbcrnjrkj6

oh no

next tuesday i am speaking at @aiDotEngineer code summit where i teach for 60 min how to build a harness from an empty loop. here's the blog post before the talk.

tej.as/blog/how-to-bu…

It's easy to get an AI agent to execute a task.

The challenge is getting them to do it RELIABLY (here's what to do)...

Once your agent loop works, there are 4 critical decisions you MUST make about your harness according to @TejasKumar_ .

And getting them wrong can lead to ended actions
• Leaking credentials
• Reporting success when they've actually failed.

Here are the 4 decisions:

1. Who decides when the task is complete?

Don't blindly trust the model when it says it's done.

The harness should verify the outcome against the expected result, ideally using an independent source of truth.

2. Who controls sensitive actions?

Keep:

• Credentials
• Irreversible operations
• Deterministic security checks

... outside of the model's control.

The harness should enforce where and how these actions can happen.

3. Who decides when to stop?

Separate your guardrails into two categories:

1. Budgets that end a run when limits are reached.
2. Policies that block unsafe actions while allowing the agent to continue.

4. Who's responsible when something goes wrong?

Find the first incorrect step and inspect what the model saw.

If the context was missing or misleading, fix the harness.

If the context was correct, replay the step and measure how consistently the model fails.

The takeaway?

A reliable agent means engineering a harness that can verify outcomes, enforce boundaries, and help you understand failures.

@TejasKumar_ recently broke down all of these 4 decisions, with practical examples and code, in Decoding AI Magazine.

Read it here: decodingai.com/p/4-key-decisi…

'evals are fuzzy unit tests'

can highly recommend @TejasKumar_'s talk from @aiDotEngineer

a fun run through everything evals youtube.com/watch?v=NdrSPm…

days since my mind was last blown by ai: 0

American Learns About Freedom In Germany!😂

The “moat” is shorthand for scarcity, and the only thing left to give (and protect!) is your own self of identity. love. your voice. hope. courage, packaged by the word. people are just so desperate to hear a real human, and not be take advantage of.

in a world of grifters, you can be Michael Scott