Medium

Guillermo Baigorria

Guillermo Baigorria

·

5d ago

How we fix an AI agent’s mistakes: data first, prompt last

A wrong card is rarely a prompt problem, so read what the model received before you touch a word.

A boy and a cybernetic robot dog weave through a line of striped test poles in a ranch paddock at sunset.

Guillermo Baigorria

·

Nov 3

One voice note, three requests: how we built an LLM router

Done with the Novak job, start the school job at 9, remind me at 8: how one hurried voice note becomes three pieces of work.

Illustration of a boy in a fenced field at sunset pointing up while a friendly white robot dog in an orange harness rises on its hind legs

Guillermo Baigorria

·

Oct 30

Prompt caching: same agent, same rules, about a quarter of the bill

Our test agent’s prompt looked fine and read nothing from the cache. Four moves later, it read up to 94%.

A line rising from 0% to 94%: the share of a small recipe agent’s prompt read from the cache, before and after four moves.

In

Generative AI

by

Guillermo Baigorria

·

Oct 29

What Google’s Harness² shows about improving an AI agent

The gains came from scripts and skills, not the prompt. And letting the model rewrite its harness blindly made it worse.

Bar chart: removing the skills from the improved Harness2 harness costs 5.3 points, the scripts 3.7 and the prompt-related components 1.5.

Guillermo Baigorria

·

Oct 28

Rent the model, own the harness: how to build a model-agnostic AI agent

Models change every few months. Four practices that keep an agent working whatever model sits behind it.

Rent the model, own the harness: how to build a model-agnostic AI agent

Guillermo Baigorria

Guillermo Baigorria

Head of Technology at Werk (getwerk.ai). Building an AI harness for teams that work away from a desk. Writing on agents, LLM routing and testing.

Following

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech