Daniel Arroyo

Founder & CTO @NieveConsulting. Previously built @AstroPrint3D. Sharing lessons on AI agents, engineering trade-offs and shipping products.

[Daniel Arroyo @CoDanny]

Anthropic is donating Agent Skills to the AAIF.

My take: package one repeatable workflow, then test it across agents. A portable format still needs portable behavior.

Your operating knowledge should outlast your vendor choice.

[Daniel Arroyo @CoDanny]

This is something I have been looking for a long time. A Free CLI based Mac cleanup tool:

https://t.co/HNR9yrt6tL

[Daniel Arroyo @CoDanny]

Before building agent infrastructure, ask:

Does custom code create a customer advantage?
Which requirement rules out existing tools?
Who maintains it in six months?

Include permissions, tracing, evals, and recovery in the estimate.

[Daniel Arroyo @CoDanny]

The most interesting thing about @OpenAI's Decisions API isn't the API.

It's that they responded to @typesafeai 's Jev in just 2 weeks, apparently without training a new model.

They adapted Luna instead.

Sometimes the moat isn't the model. It's how fast you can repurpose one.

[Daniel Arroyo @CoDanny]

Is it me or is @github being down a daily occurrence?

[Daniel Arroyo @CoDanny]

Try this agent test:

Give it a budget. Change the budget halfway through a long task. Ask it to act at the end.

Does the final action respect the latest constraint?

Which instruction does your agent tend to forget first?

[Daniel Arroyo @CoDanny]

I'm bullish on WebMCP.

OpenAI already supports it through ChatGPT desktop site tools. Websites can expose structured actions for AI agents to use.

I'm going to start adding it to my websites.

Still early, but if you build for the web, take a look.

[Daniel Arroyo @CoDanny]

Give each piece of agent context a job.

Needed now? Keep it active.
Needed later? Make it retrievable.
Useful next week? Consider persistent memory.

Keep the original evidence accessible. Summaries can lose the detail your next step needs.

[Daniel Arroyo @CoDanny]

A booking agent can write a perfect confirmation for the wrong date.

Evaluate:
- The user's constraints
- The tools and arguments
- The actual booking record

The final message is only one part of the result.

[Daniel Arroyo @CoDanny]

That said, I'm very bullish on this direction.

The future isn't one giant LLM doing everything.

It's powerful reasoning models + decision models + deterministic code.

I want Jev to succeed. I genuinely hope I'll have a reason to adopt it soon.
https://t.co/Tt8fSQNmJh

[Daniel Arroyo @CoDanny]

Hot take: We're wasting ridiculous amounts of AI compute asking general-purpose LLMs to make simple decisions.

@typesafeai 's Jev might be one of the most interesting attempts to fix this.

I seriously considered using it in my agents.
Then decided not to.
Here's why 🧵

[Daniel Arroyo @CoDanny]

And there's the dependency question.

A new provider means another API, another potential failure point, and another system to maintain.
I'm happy to introduce dependencies when they solve real problems.

I'm not interested in collecting vendors because their technology is cool.

[Daniel Arroyo @CoDanny]

If compacting your AI agent’s history changes what it’s allowed to do, you have an architecture problem.

Keep state durable, evidence traceable and memory disposable.

https://t.co/NXXVUnBeFI

[Daniel Arroyo @CoDanny]

Start your agent evals with one costly failure.

Pick a behavior. Collect realistic cases. Record a baseline. Add failures as you find them.

Check the actions and the final state, too. A convincing "Done!" is a very cheap acceptance test.

[Daniel Arroyo @CoDanny]

Your next AI feature may already be in your support queue.

Find a recurring question. Test the workflow with staff. Turn failures into eval cases. Preserve user permissions. Expand when the evidence holds up.

Which question keeps returning?