Ten items are on today’s radar, and one clear thread runs through almost all of them: what used to be an engineering project is becoming an off-the-shelf offering. I started with the launch that changed my priority list the most.

In brief

Presence, OpenAI’s new platform for putting real-time voice agents and chatbots into operation, changes the question for product teams: instead of estimating only how long it takes to build, it is worth asking whether building still makes sense. In financial services, buying the voice layer does not replace the design of rules, authorization, interaction records and human escalation. That is where the most important part of the product remains.

Anyone who works with credit products knows that much of the friction is not in the analysis. It is in the conversation.

The customer calls to understand an advance payment. They send an email to ask why an invoice was not accepted. They wait for an answer about a limit. Each of these points is a queue, an SLA and, ultimately, a cost.

That is why the launch of Presence caught my attention. It is OpenAI’s new platform for companies to put real-time voice agents and chatbots into operation. It is still in limited availability, with deployment supported by OpenAI’s own engineers and integration partners.

What changes when the voice layer comes ready-made

What changed here is not the voice technology. It is the format of the offering.

Until recently, a company that wanted an intelligent service agent would buy an API and assemble everything in-house: orchestration, context, integration and governance. Now there is a ready-made option, with an implementation team alongside it. For people on the product side, this changes the conversation from “how long will the team take to build it?” to “does it make sense to build it?”

This is a central decision in AI product management: separating infrastructure that can become a commodity from the logic that differentiates the business. The ready-made layer can accelerate the path to operation, but it does not answer on its own how the product should behave.

What remains the product team’s responsibility

And it is not an obvious decision. In financial services, the agent needs to know what it can and cannot say, record every interaction, respect authorization limits and escalate to a human at the right time. None of that is a detail; it is the product itself.

Buying the voice layer does not eliminate the work of designing those rules. It only shifts the effort to where it really matters: the design of AI agents, AI governance controls and limits for each action. Teams evaluating how to create agents can use this guide to building AI agents to structure that decision before implementation.

This is the kind of movement that excites me. The laborious part becomes a commodity, leaving more time to think about what only people who know the business can solve.

For anyone who is curious and wants to read the full story, here is the link: OpenAI unveils Presence.

The rest of the radar

Kimi-K3 on Hugging Face — redefines the ceiling for cost and capability in open models, enabling features that previously ran only on proprietary APIs. Read more

Dynamic Workflows in Claude Code — the agent assembles its sequence of steps at runtime, shortening the path from discovery to a functional MVP. Read more

Robinhood lets AI agents trade stocks — the first relevant case of an agent taking irreversible financial action: a reference for consent, limits and auditability. Read more

HubSpot opens Agent Hub and Agent Builder in public beta — GTM teams can create agents without engineering, and the discussion becomes who defines permission and quality. Read more

Noisy LLM evaluators still improve agents — challenges the excuse that “we do not have a reliable eval”: teams can iterate with an imperfect metric and measure real progress. Read more

Step 3.7 Flash — another fast, low-cost option for high-volume features, where latency and cost per call decide the outcome. Read more

AISlop, a CLI for detecting AI code smells — points to a new kind of technical debt that needs to enter the conversation about team quality and capacity. Read more

Study maps dark patterns in AI chatbots — optimizing session time in a conversational product can produce manipulation by accident: a brand and regulatory risk. Read more

DX Core 4 for measuring engineering productivity — a defensible framework for answering whether AI adoption increased delivery or merely shifted effort to review. Read more

That is it for today. If “let’s build this” comes up today, ask first whether someone is already selling it ready-made.