There were plenty of model launches in today’s radar, but something else held my attention: the moment an AI agent stops answering questions and starts planning its own work process. That is where I want to begin.

Working in credit products, I often ask myself: how far can we delegate real work to an AI agent before it becomes nothing more than a brainstorming assistant?

Anthropic has just launched “dynamic workflows” in Claude Code. In practice, Claude itself can write the orchestration scripts that put dozens—sometimes hundreds—of subagents to work in parallel within a single session, then review its own work before delivering it. These workflows have already been used for bug audits, code optimization, and large-scale migrations; in one case, roughly 750,000 lines were ported in 11 days.

What stands out to me is not the line count. It is the change in role. Previously, an agent answered one question at a time. Now it plans its own work process and delegates tasks to itself, almost like a team.

That directly changes the scope of a roadmap. Migration, pipeline automation, and receivables reconciliation—work that now takes months because people need to review every line—can start to become a matter of days, with humans stepping in only for final validation.

This is not about taking people out of the process. It is about where people contribute: less repetitive execution, more risk decisions and judgment about what truly matters. For financial products, that is an opportunity, but it also requires careful design of the human approval point.

For anyone who wants a closer look at the launch, the article is here.

The rest of the radar

Kimi K3, a new frontier-level open model — Another strong open-weight option puts pressure on the pricing and differentiation of LLM-based products. Read more

NotebookLM became Gemini Notebook — It shows Google’s strategy for brand unification and integration across its AI products, a pattern worth watching. Read more

LM Studio launches Bionic, an agent for open models — It reinforces the trend toward local, privacy-first agents as an alternative to cloud-based agents, relevant to data-sensitive products. Read more

“Noisy” AI evaluators can still help improve agents — It offers practical grounding for deciding when costly human evaluation is justified versus a cheaper LLM-as-judge in an agent improvement cycle. Read more

Robinhood lets AI agents trade stocks — A concrete case of agentic commerce in a regulated product, with permission patterns and a dedicated portfolio that can serve as design references. Read more

Step 3.7 Flash, an affordable model built for agents — It expands the cost-performance options for products with heavy AI use in coding and search. Read more

Study maps “dark patterns” in AI chatbots — A direct reference for anyone defining ethical limits for engagement and retention in conversational products. Read more

Promptloop, a lightweight tool for testing prompts in the terminal — It makes it easier to validate AI feature changes before production, useful for lean teams. Read more

Claude Fable 5 vs. GPT-5.6 Sol in a $100 music video — A practical benchmark of video-model cost and quality, useful for model-sourcing decisions. Read more

It was a day full of launches and little consensus on where all of this is headed. I will keep tracking it and bring more tomorrow.