τileo

Building AI agents that do real operational work. AI advisory + managed agents for relationship-heavy businesses.
Writing about AI agent memory and judgment.

[τileo @Tileo_LG]

Most people building AI agents ask: “How do I make the agent smarter?”
I think the better question is: “What should this agent be allowed to remember?”

[τileo @Tileo_LG]

@sama You have a Penalty Shot !

[τileo @Tileo_LG]

Before building an agent, split the workflow into three buckets. Keep human judgment with people. Automate only memory and context. Repeat and record outputs until you have a pattern and something measurable to improve iteratively.

[τileo @Tileo_LG]

@Teknium It's working but i feel it way more slow than before the last update.

[τileo @Tileo_LG]

That’s not judgment.
That’s operational drag.

[τileo @Tileo_LG]

Nobody serious wants a zero-human company.
They want the human to stop doing the work that doesn’t require being human.

[τileo @Tileo_LG]

Remembering follow-ups.
Chasing notes.
Rebuilding context.
Checking if someone replied.

[τileo @Tileo_LG]

@gregisenberg So much truth in the last point. Ahead to be hear but not too much to be understood. I think that will be the roles of the tomorrow consultant. Being able to understand a customer situation, experience with tech and adjust the solution accordingly to their need and capacity today

[τileo @Tileo_LG]

@dotta Best pivot, I really believe no one want zero human company, we just want best ouput per human hour.

[τileo @Tileo_LG]

@mamagnus00 Agency

[τileo @Tileo_LG]

Today I relearned a boring but important AI agent lesson:

A green check can still be wrong.

I saw a dashboard domain marked as misconfigured.
The obvious fix was to point DNS to the provider’s recommended target.

It was technically correct.
But operationally wrong.

That subdomain was intentionally routed through a tunnel to a customer instance.

Lesson:

Agents don’t just need API access.
They need a map of intentional exceptions and decisions already taken.

Without that, they optimize the metric and break the system.

[τileo @Tileo_LG]

@itsolelehmann what about browser-use compared to Browserbase?

[τileo @Tileo_LG]

@garrytan Thank you for sharing this.
It genuinely opened my mind.
I’ve been building MAIDA (Managed AI Digital Agents) for consulting boutiques exactly the kind of teams that need to duplicate their expertise without hiring 10 more people.
Your architecture resonated deeply:

- Thin/light harness (I’m using Mainly Hermes)
- Two separate GBrains (one Company brain + one Individual brain) so the agent has both shared operational context and personal founder judgment
- A Skillify-equivalent meta-skill that extracts repeatable patterns into reusable, versioned skills
- A clear evaluation + optimization loop for every skill we generate
- A second optimization loop curated by a human operator in the loop.

The one thing I’ve added for customer reliability (especially while the system is still learning) is a webhook + deterministic toolkit layer.
It lets us inject guardrails and structured processes on top of the agentic flow. I suspect this becomes less necessary as the brain accumulates enough usage patterns but right now it gives my clients the confidence to actually use it in real client work.

This whole piece made me realize I’m not just building agents. I’m building a compounding nervous system for small professional services teams.

Grateful for the clarity. Going to reread it a few more times.

[τileo @Tileo_LG]

@NousResearch @OpenRouter Congrats, and most important for me you are the first agent framework reliable enough to be integrated on customer projects. You change Personal Agent from nice side project and experience to service sellable and manageable in a lot of industry.

[τileo @Tileo_LG]

@nathangotch You missed X

[τileo @Tileo_LG]

As an Industrial Engineer who’s spent 10+ years moving designs from CAD to Real production, I’m totally impressed by that.

The dexterity is on another level, but what really impressed me is the full-stack architecture. Most robotics projects fail because they add AI on top of existing hardware. Genesis choose the hard route and it’s clearly paying off.

For years, the bottleneck wasn’t intelligence. It was reliable, contact-rich manipulation learned from real humans at scale.
They just solved a big piece of that.
This changes what’s possible for small teams. At AI Jungle we’re already building MAIDA (Managed AI Digital Agents) to help consulting boutiques duplicate their expertise.

Next step: MAIRA Managed AI Robot Agents that let boutiques prototype and bring ideas to life in-house using these new capable robots + additive manufacturing, without needing a robotics and AI team in house.

The future of fast physical iteration just got a lot closer. and with fast iteration you know how it will compound. So excited to live in today world. Thank you @gs_ai_ and congratulations.

[Genesis AI @gs_ai_]

We are back. After one year of quiet building.

Introducing GENE-26.5, our first robotic brain that takes a major step toward human-level capability.

For years, robotics has struggled to learn from the world’s largest and valuable data source: Humans.

Solving it means rethinking the whole stack from the ground up:

- A robotics-native foundation model.
- A 1:1 human-like robotic hand.
- A noninvasive data collection glove for motion, force, and touch.
- A simulator that turns weeks of experiments into minutes.

GENE-26.5 is trained across language, vision, proprioception, tactile, and action. We designed a set of tasks to test how far we can go with this new paradigm.

Fully autonomous, 1x speed, one model, same weights. (Enjoy with sound on)

We are approaching the endgame for robotics.

And this is just a beginning.

[τileo @Tileo_LG]

@donk8r Sell output and service to nework helps me for first revenue, and customer feedback. Soon it will be time to productize it for every one

[τileo @Tileo_LG]

Bug I fixed today:
The AI agent “remembered” the user’s preferred output format.

But the scheduled job that actually sent the message was just deterministic Python.

So it ignored the memory and shipped the old weak format.

Lesson:

If your product has agents + scripts + crons, memory is not enough.

You need preference propagation across the whole execution path.

Otherwise the agent sounds smart in chat and dumb in production.

[τileo @Tileo_LG]

@eliana_jordan would have said first last month but i am really surprised by my Latest A/B testing, looks like reply is stronger for not formatted message. happy to read about your experience, if you had run them on some test subjects.

[τileo @Tileo_LG]

@icanvardar I am taking my car to wash it at the carwash and eating strawberries with 3 r.

[τileo @Tileo_LG]

@floriandarroman Same kind of experience few weeks back. Looks like a second personal agent honeymoon. Enjoy and ship