Agentic Engineering / Published from the workbench
We study how capable models become accountable products: the harness, router, control plane, evidence, human decisions, and production consequences around them.
Weekly substantial note
Biweekly email
Complete English + Spanish editions
Archive
13 issues · no filler
Nine small MIT-licensed skills that make an agent's claims checkable: what was checked, against which code, and who handed off what.
A persisted lease can recover a background task without making an external effect exactly-once. A 50-run fault-injection test separates task ownership, HTTP requests, and accepted effects.
A robot, a station and reversible gravity. The result is playable; the process shows an agent building, testing and correcting under supervision.
We compared four versions of a fictional order. Two extractors recovered a date and an instruction that did not appear on the rendered page. The result separates three things an agent should not conflate: reading, substantiating and authorizing.
Dots offers a way to delegate work that can continue between conversations. For a team, a useful starting point needs a bounded goal, clear permissions, and a result people can review.
A specialized model can resolve a classification without generating a conversation. The opportunity is to simplify that part of the system, not to confuse a valid response with a reliable decision.
What one anonymized, multi-slice regulated-industry push taught us about turning Orca, OMP, heterogeneous models, issue trackers, and CI into one accountable delivery system—and why adding agents is the easy part.
Eleven models each received two recorded runs on the same ten animated 3D tasks. Four judges scored the selected artifacts. The useful result is not a universal winner. It is a record of what happened when models, output channels, schema validation, a shared renderer, frame capture, and a model panel were treated as one executable system.
On 20 July 2026, Claude Fable 5 became included in Max and Team Premium plans, capped at half the weekly usage pool. Four days later, Claude Opus 5 arrived on the API at half Fable 5's rate. What one week did to the economics of verification, and where a subscription lane must never carry a production dependency.
Five models and one access change landed between 16 and 24 July 2026. This is not a roundup. It is a routing note: which lanes got re-priced, which access boundary moved, what three of the four vendors did not publish in checkable form, and the policy we run because of it.
In July 2026 an internal OpenAI capability evaluation reached Hugging Face's production systems. Both companies have published accounts, and the accounts are not yet reconciled. The lesson does not depend on which one holds: the boundary around adversarial AI testing is a third-party risk surface, and it deserves production-grade controls. Written to the public record as of 25 July 2026.
A field guide to expanding an agent's authority through bounded action spaces, exact approvals, failure drills, canary exposure, and recovery that works before production depends on it.
A field guide to putting GPT-5.6 or Fable 5 inside an accountable operating system: one that routes deliberately, acts through a harness, keeps receipts, and ships work humans can verify.
Weekly signal · biweekly email
One substantial note when we have something worth showing: systems, receipts, prompts, and what failed. Confirm by email. Unsubscribe whenever you like.
