Jaime Vengoechea
AI-powered. Shaped by Jaime Vengoechea.

The Applied Future.

AI, work, technology, and what comes next.

Sep 12, 2026Source-linked briefingPrevious editions RSS feed
In this edition

Six fresh selections connect cloud agents, evidence-based software review, marketing measurement and better scientific data. The practical question is no longer just what AI can generate, but how to check whether it helped. These reports were published September 10-11; their original dates are preserved. Earlier selections remain in the seven-day reading list and dated archives.

This Edition

Sep 12, 2026 · 6 stories
Today's leadPublic beta announcement

OpenAI opens its Codex agent infrastructure to developers

OpenAI introduced the Agents API in public beta, giving developers access to its managed Codex harness for long-running cloud agents. It supports tools, context management and subagents, with a choice of OpenAI-hosted or other supported environments. The announcement describes infrastructure for building workflows, not a guarantee that an agent will complete every task correctly.

Read original source
Context & source check

For builders, managed agent infrastructure could reduce orchestration work. Keep permissions narrow, define a verifiable deliverable and compare total operating costs before moving a recurring workflow into production.

Public beta announcement. Company source. Source checked Sep 12, 2026. Not independently tested.

The reading list

5 stories

More from This Edition

5 stories
Latest editionCompany-published customer case study

Devin's next deliverable is evidence that the software works

An OpenAI case study describes Cognition using GPT-6 Astra inside Devin to test applications and return evidence. Its example includes a simulator recording and a report separating passed checks from untested areas. Cognition presents this as a path toward more efficient review, rather than proof that engineers can stop reviewing software.

Read original source
Context & source check

Ask an AI builder for a working demo, test results and explicit gaps alongside the code. Evidence that a workflow behaves correctly is more useful than a confident completion message.

Company-published customer case study. Company source. Source checked Sep 12, 2026. Not independently tested.

Image: Source: Google. Google Ads is a trademark of Google LLC; no affiliation or endorsement. Image source & credit.
Latest editionMeasurement product announcement

Google connects first-party data with more rigorous marketing measurement

Google announced Data Manager integrations with Analytics and Display & Video 360, a Data Strength Uplift metric and upgrades to Meridian, its open-source marketing mix model. Meridian GeoX is now generally available globally for geo-experiments. The package combines better data connections with tools for investigating incremental impact; vendor-reported conversion uplift is not a promised result for every advertiser.

Read original source
Context & source check

Treat recovered conversion reporting and genuinely incremental demand as different questions. Audit your data connections, then use experiments to test whether a campaign changes business outcomes.

Measurement product announcement. Company source. Source checked Sep 12, 2026. Not independently tested.

Latest editionTechnical implementation walkthrough

An agent can stay online and still fail the user

AWS demonstrates a two-layer monitoring approach: AgentCore Evaluations assesses agent quality while AWS DevOps Agent investigates infrastructure incidents. Its airline-reservation example distinguishes successful tool calls from actual task completion. The authors also warn that model-based evaluation scores need human calibration and that sampled, after-the-fact checks cannot prevent every problematic response.

Read original source
Context & source check

Monitor both whether your automation ran and whether the intended result happened. Add an end-to-end completion check, a failure alert and human review for consequential outputs.

Technical implementation walkthrough. Company source. Source checked Sep 12, 2026. Not independently tested.

Image: Source: Google. Gemini is a trademark of Google LLC; no affiliation or endorsement. Image source & credit.
Latest editionDesktop app launch

Gemini moves into the Windows desktop workflow

Google launched a Gemini app for Windows 10 and 11, with an Alt + Space shortcut for help over the current task. The app also supports drafting with connected Google apps and creative tools. Google lists global desktop availability, while some agent and video features require an AI subscription, have availability limits and are restricted to adults.

Read original source
Context & source check

Desktop access can make an existing tool easier to use. Try one repeatable task, check connected-data permissions and verify which features your account actually includes before changing the team's workflow.

Desktop app launch. Company source. Source checked Sep 12, 2026. Not independently tested.

Euclid telescope data examples with blue masks marking possible image artifacts
Image: ESA/Euclid Q1 Data release; image processing by Aimee Schechter and Bharath C. Nagam, via NASA. Image source & credit.
Latest editionCitizen-science project

Help astronomers teach AI which telescope signals to ignore

NASA's Artifact InSPECtor invites volunteers to identify misleading artifacts in space-telescope data and help improve AI-assisted cleaning. Participants work with Euclid observations; Roman telescope data is expected to join in early 2027. The project addresses errors caused by instruments and stray signals before astronomers use the observations to study galaxies and the universe.

Read original source
Context & source check

Human feedback is part of the system, not an afterthought. This is also a practical way to participate in science using a phone or computer, without needing to build a model yourself.

Citizen-science project. Research institution. Source checked Sep 12, 2026. Not independently tested.

Story preview