AI / agent engineer. Geophysicist by training, federal public servant by day (ANP, posted at ANM), and for about six months now the product owner of a fleet of coding agents.

I do not architect or author the code. Agents do — worker seats produce ~95% of it, per our own published engineering standard (harness-standards/standards.md §1). I author the intent, ratify the plan and the validation contract, and own the risk. So the engineering I actually do is in the contracts, the gates and the evals around the agents, and that is what these repositories are evidence of — including where it went wrong.

• Validation contracts that can fail. A mission is not done because a seat exited 0; it is done when a held-out validator checks the diff against a ## Done when section that names a mechanism which executes. The 68-entry decision ledger behind that claim is public, and five of its entries correct earlier entries in the same file — see the essay and decisions.md in agent-factory.
• Review surfaces that refuse to flatter. eval-viewer carries a regression suite with one block per false claim the dashboard has actually made — a route credited to a node that cannot reply, a green check on a run that recorded no verdict. Those blocks are not deleted to make a change land.
• Gates that report honestly when their tool is off. A metric whose linter rule was never enabled read 0 for weeks and the baseline pinned it as a fact; enabling it surfaced 74 real violations. The question every new check now has to answer is what it prints when it is measuring nothing. The second essay is about that.

These are working artefacts from a small business, not a portfolio exercise. They have retired deployments, absent files and open debts, and the READMEs say which.

• Six months of agent-built software: what 68 decisions say — a reading of an append-only decision ledger, concentrating on the entries where the process was wrong and said so.
• An eval gate on real WhatsApp traffic — why a check that cannot fail is worse than no check, and what that costs to fix.

~20 years writing code, 15 of them Python, at the intersection of physical science and high-performance computing: five years at Schlumberger on production geoscience software, then a decade-plus automating technical and GIS workflows inside Brazilian energy and mining regulation. Contributor to Fatiando a Terra since 2012, plus GeographicLib and geonumerics. On Stack Overflow as imbr: 7.9k reputation, 85 answers over 14 years, 5.4M people reached, top answer at 197 votes; NLP Collective member.

The regulatory and scientific-computing background is not decoration here. Most of what agent engineering turns out to need — evidence, audit trails, and a definition of "done" written before the work starts — is the same discipline, applied to a faster machine.

eusoubrasileiro@gmail.com · GitHub · Stack Overflow · LinkedIn