Seroter's Daily Reading — #766 (April 17, 2026) — Seroter's Daily Reading

Listen: https://blossom.nostr.xyz/d5b8a49ed876e8852288e049e57afd733c859c6c942806920ed5bdb25520f77f.mpga

Seroter's Daily Reading, Episode 766 — April 17, 2026.

Big news this week, and a lot of it clusters around the same theme: how established platforms are repositioning themselves to survive in a world where AI agents are the primary users, not just helpers. Let's get into it.

Leading this off with Salesforce's Headless 360 announcement, which made its most ambitious architectural move in the company's 27-year history this week at its TDX conference. More than a hundred new tools and skills shipped immediately. The company even went so far as to say it rebuilt Salesforce for agents over the past two and a half years. This is a company looking at the enterprise software sell-off, at the fear that AI might render traditional SaaS obsolete, and essentially betting that if agents are coming for your job, the best move is to become the best possible substrate for those agents to operate on. The key architectural insight from the interview they gave VentureBeat is the distinction between two agent patterns they're supporting: customer-facing agents that demand tight deterministic control with explicit business rules, versus employee-facing agents running what one executive colorfully called the Ralph Wiggum loop, where the agent dynamically decides its own path at runtime. Both patterns run on the same platform. That's the unifying bet. Also notable: the company is moving from per-seat licensing to consumption-based pricing for Agentforce, which is a quiet but massive business model shift. When agents, not humans, are doing the work, charging per user stops making sense.

Over at Google, AI Mode in Chrome desktop now opens webpages side-by-side with the AI Mode panel, so you can search, click a link, and keep the conversation going without tab hopping. A nice quality-of-life upgrade if you're already living in Search. The reaction online seemed positive, which is not always the case when Google ships something.

On the topic of platforms choosing not to force AI on users, there's a piece on Canva's AI 2.0 announcement with Melanie Perkins making a point worth dwelling on. But here's the thing: you can ignore it entirely and use the platform exactly the way you always have. That's a genuinely different stance from most large software companies right now, who are essentially redesigning their UIs around AI whether you asked for it or not. Canva has more than a quarter billion monthly active users, most of them non-technical, and they're betting that optional is the right call at their scale. Perkins also mentioned they trained their own foundation model, with more than a hundred people in the research team, and hit a key breakthrough in October around generating designs in fully editable layered formats rather than flat images. Interesting also that Canva has been profitable for nine years, which means the CEO isn't on a public earnings call trying to explain away AI disruption.

The A2A protocol turned one year old this week. Google published a retrospective, and it's worth reading because the momentum has been real. The protocol went from a Google announcement to a Linux Foundation donation within two months, and now has over a hundred technology companies supporting it. They hit version 1.0 in March, which is a real milestone. The community also shipped signed agent cards for cryptographic identity verification and refined the web-aligned architecture for enterprise deployments. A2A is designed to be complementary to MCP, which handles tool integration, while A2A handles the coordination between autonomous agents. The Google post also flags that the A2Family is growing, with protocols like AP2 for payments, A2UI for user interfaces, and UCP for commerce, all built on the same open extensibility model. There's even an A2April celebration happening this month, if you're into that sort of thing.

On a related note, A2UI v0.9 shipped this week too. This is the framework-agnostic standard for what Google calls generative UI, the ability for agents to dynamically generate interfaces that match your existing design system on the fly. The big change in this version: frontend developers don't want new components, they want their existing design system, so they renamed the optional component set from Standard to Basic to make that clearer. They also shipped an official React renderer, a shared web-core library, and a new Agent SDK that makes integrating A2UI into any Python agent as simple as a pip install. The ecosystem is growing too. AG2, the creators of AutoGen, built native A2UI support. Oracle shipped support across Agent Spec, AG-UI, and A2UI. Vercel has a json-renderer that supports A2UI. There are live demos now for health and financial planning applications, and you can get started with a single npx command.

Shifting to product thinking, there's a really good piece this week from Marty Cagan's team at SVPG on the distinction between building to learn and building to earn. This is a framing coined years ago by product coach Jeff Patton, but it resonates especially strongly in the age of generative AI. The idea is that product discovery is fundamentally about building prototypes to learn whether you're solving the right problem, while product delivery is about building something commercial quality you can sell, service, and support. What's changed now is that the cost of building has dropped so dramatically that delivery is no longer the bottleneck. The real bottleneck is in discovering a solution worth building in the first place. Cagan's team argues that too many product teams have simply become turbo-charged feature factories, capable of producing bad products faster than ever. The best product managers are leaning into build-to-learn skills and developing what they call product sense, and that's where the real value is for people working in product right now.

Speaking of agents and human collaboration, there's a new empirical study from the University of Saskatchewan that looked at nearly ten thousand agentic pull requests on GitHub across more than a thousand repositories. The researchers divided developers into core contributors, those in the top twentieth percentile of contribution experience in their repo, and peripheral contributors. Both groups used coding agents at similar median rates, but the peripheral developers had a subset who were heavy power users of agents, delegating broadly across bug fixes, features, docs, and tests. Core developers concentrated agent use on documentation and testing, the tedious work. More interesting: core developers reviewed agent output more intensely, with higher comment counts and more design-level feedback, while peripheral developers more often just caught bugs. And peripheral developers were nearly twice as likely to skip CI checks before merging. The researchers' advice: establish consistent review standards for agentic PRs regardless of who delegated the work, because the social layer and project familiarity still govern quality gates even when the code was written by an agent.

Gemma 4 is the subject of a detailed post from NVIDIA this week, their open multimodal model family. The bundle includes four models: a 31B dense transformer, a 26B Mixture of Experts model with 128 experts, and two smaller on-device variants. All support text, audio, vision, and video, and all are available under the Apache 2.0 license. The 31B and 26B A4B support 256K context windows. The smaller E4B and E2B variants target edge and mobile deployment, including Jetson Orin Nano. NVIDIA worked with vLLM, Ollama, llama.cpp, and Unsloth for deployment across DGX Spark, Jetson, and RTX platforms. There are getting-started guides for each. Enterprise users can also access the 31B model via NVIDIA's hosted NIM API for free prototyping.

From HR Dive, a report from Docebo that is worth sitting with. Companies are spending heavily on AI readiness training, but eighty-five percent of employees say they can't apply the training they received to their day-to-day jobs. Fifty-six percent of workers are so overwhelmed by manual pre-AI tasks that they don't have time to learn the tools that are supposed to save them time. Seventy-eight percent of learning happens outside the tools they actually use, like Slack or Salesforce, which makes the training feel like a distraction rather than an investment. The experts quoted in the piece suggest rethinking this with a timeline rather than a one-shot approach, getting employees collaborating in spaces like Slack channels for AI wins, and starting with the actual problem you had before you throw AI at it. The broader point: AI upskilling that isn't applied within two weeks is mostly forgotten, and companies are discovering that the change management problem is much harder than the technology problem.

There's also a piece titled Sleep Deprived from Robert Glazer that ran this week. Glazer walks through some of the sleep research, including a landmark Oxford study that found people sleeping six hours a night performed as poorly as someone who hadn't slept at all in 48 hours by day fourteen. The more troubling finding: subjects stopped noticing their own cognitive decline after just a few days. Their self-reported sleepiness leveled off even as their actual performance continued to deteriorate. Glazer applies this as a broader metaphor too. Slow erosion is the most dangerous kind because we convince ourselves it's not happening. It's a good reminder to close the episode on.

And finally, from Guillaume Laforge, a hands-on walkthrough of Gemini 3.1 Flash's new expressive TTS model, this time with a Java implementation. The new TTS model can be steered with audio profiles, scene descriptions, and director's notes, and you can use inline tags like excitedly or whispering to change emotional delivery mid-sentence. The demo sets up a Morning DJ persona and streams the audio directly to the speakers as it's generated, using the Gemini Interactions SDK for Java. If you've been looking for a way to get expressive TTS into a Java application, this is a solid starting point.

That's episode 766. Next episode coming next week. I'll be at Google Cloud Next in Las Vegas, so if you're there, stop by the booth and say hello.

Salesforce launches Headless 360 to turn its entire platform into infrastructure for AI agents — VentureBeat
A new way to explore the web with AI Mode in Chrome — Google Blog
Canva won't make you use its new AI — Sources
A year of open collaboration: Celebrating the anniversary of A2A — Google Open Source Blog
Build to Learn vs Build to Earn — SVPG
A2UI v0.9: The New Standard for Portable, Framework-Agnostic Generative UI — Google Developers Blog
How does codebase familiarity shape collaboration with coding agents? — Research-Driven Engineering Leadership
Bringing AI Closer to the Edge and On-Device with Gemma 4 — NVIDIA Developer Blog
Why AI readiness training fails — HR Dive
Sleep Deprived — Friday Forward
Streaming Gemini 3.1's expressive new TTS model in Java — Guillaume Laforge