Ask before acting
Agents write drafts, not data. A person presses Confirm, and the action runs with that person's own permissions.
Agents online · illustrative replay
Founder of Defabric and MNDA Estudio. I design agentic workflows on Claude — tool scoping, approval gates and cost guardrails — and ship them to the plant floor and the back office. Based in Mexico.

Re-plan Extrusion 1
The agent proposed a draft. Nothing is written until a person confirms it, with their token and their permissions.
Illustrative diagram: a person asks; the Director agent delegates to the Production lead, who delegates to the Extrusion coordinator; it reads what is pending and proposes a schedule as a draft; the draft goes back to the person, who confirms or rejects it. Only on Confirm is anything written, with that person's permissions.
Three projects running today. Each case explains the architecture, what the agent cannot do without a person, and what is measured (and what isn't).
A Director agent delegates to department leads and work-center coordinators. Every write waits for a person to confirm it. In production at 2 factories.
Read the case
Content, web/apps and music pipelines, with approval gates the code enforces, spend caps and request tracking.
Read the caseNot principles on a slide: they're checks in the server and in the scripts, the same in Defabric and in MNDA.
Agents write drafts, not data. A person presses Confirm, and the action runs with that person's own permissions.
Each agent only gets the tools its team needs, intersected with what the user is allowed to do.
Token usage per user and per agent; daily and per-job spend caps the scripts refuse to exceed.
AI-authored automations that then run without AI. Screen reads in 28 ms (median) vs ~2.5 s with uiautomator dump.

Shared expenses; baby care with a breast-milk bank.


Email is the fastest way to reach me.