Diary / Field Notes
This is the fast lane: short observations, small updates, early lessons and things worth noticing before they become full essays or project notes.
Five signals from this week point in the same direction: access, security, compliance, agency economics and campaign infrastructure are now the real AI story.
The AI story is moving out of demos and into permissions, monitoring, compliance, pricing and operating design.
Estonia says it will move forward with official digital identities for AI agents, so delegated software can act with limited rights and a clear audit trail. This is agent infrastructure with legal consequences.
AI agents will not be trusted at scale until they can prove who they act for, what they are allowed to do, and who is accountable.
Ad Age reports that Accenture Song is launching Marketing Investment Navigator, using Amazon Ads data and agentic AI to analyse cross-channel media performance. The useful signal is measurement, not another dashboard.
The serious agency AI race is moving toward where money is allocated, not where copy is drafted.
WPP Media is developing a Buyer Agent for Video with Disney, Netflix, NBCUniversal, Fox, Paramount, Comcast Advertising, IAB Tech Lab and Prebid.org. The signal is not automation. It is governance for agentic media spend.
Agentic media buying is moving from demo theatre into the plumbing of real brand budgets.
Google, Microsoft, GitHub, Nvidia, Salesforce and others are backing Agentic Resource Discovery. It is plumbing, but plumbing is what turns agents from demos into working systems.
Agents do not just need better reasoning. They need a trusted way to find the right tools, skills and other agents at runtime.
Perplexity's Brain feature points at the next useful layer in agents: remembering what happened during the work, then using that memory to reduce repeated effort.
The next serious agent advantage is not a warmer personality. It is work memory that stops the system making the same mistake twice.
Adobe Brand Visibility is not another SEO dashboard. It is a signal that being cited by ChatGPT, Copilot, Perplexity and Google AI Mode is becoming a measurable marketing surface.
AI visibility is turning from a hunch into an operating metric, and that changes how brands need to think about being found.
Salesforce's $3.6B Fin deal is a simple bet: customer support is moving from ticket handling to autonomous resolution, and vertical agent performance now has platform value.
Fin shows why vertical agents are becoming acquisition targets: the value is not conversation, it is verified resolution at lower cost and higher speed.
SpaceX has agreed to acquire Anysphere, the company behind Cursor, in an all-stock deal valuing the AI coding company at $60B. This is not just Musk buying a code editor. It is a bet that software creation has become strategic infrastructure.
The Cursor deal is a loud signal that coding agents have moved from developer productivity tool to core strategic infrastructure.
Midjourney says it will announce its first hardware project at 6pm PT on June 17 from San Francisco. The details are still scarce, which is exactly why the signal is interesting.
Midjourney moving from image generation into hardware is worth watching because the next creative interface may not look like a prompt box.
GitHub Agentic Workflows is in public preview, Copilot Chat can now see past agent sessions, and the Copilot SDK is generally available. The developer stack is moving from chat prompts to persistent agent work.
The important GitHub move is persistence. Agents are leaving the chat window and becoming repeatable, searchable, policy-constrained engineering infrastructure.
Reported leaked OpenAI 2025 financials point to about $13B of revenue against about $34B of spending, including roughly $10.6B paid to Microsoft for research and development expenses. The signal is frontier AI economics, not model hype.
Frontier AI is not just a capability race. It is a capital-allocation race where revenue can look enormous and still be dwarfed by compute, training and infrastructure.
Coinbase says Coinbase Advisor is an AI-powered automated advisor offered by a registered investment adviser and commodity trading adviser. The signal is clear: agentic finance is becoming a compliance product, not just a demo.
The important Coinbase signal is not that an agent can trade. It is that agentic finance now has to be designed around permissions, limits, compliance and regulated advice.
Salesforce has signed a definitive agreement to acquire Fin, formerly Intercom, for about $3.6B. The useful signal is not the M&A headline. It is customer service moving from ticket workflow to autonomous resolution.
Salesforce paying $3.6B for Fin is a clean validation of the Foundry Works thesis: customer experience is becoming agentic, measurable and operational.
Nex AGI has put a 397B total parameter agentic MoE model on OpenRouter with a free route. The useful signal is not the headline size, it is what free agentic capacity does to the cost of experimentation.
When a 397B total parameter agentic model is free to test, the sensible move is not to worship the benchmark. It is to throw real work at it before the pricing changes.
Hermes Agent now supports xAI Grok OAuth, so a Grok subscription can power reasoning, voice and image/video generation inside a persistent open-source agent.
A consumer Grok subscription can now plug straight into a persistent open-source agent with memory and multimodal output.
A short X video is a useful signal: AI-generated UGC-style ads are moving from novelty toward something brands will actually test in the wild.
The important shift is not that AI can make video. It is that AI can now produce the kind of messy, direct, UGC-style creative that performance teams actually use.
SAP is embedding Claude as a primary reasoning and agentic capability across SAP Business AI Platform. The enterprise agent thesis is moving from pitch to product.
Agentic AI in regulated enterprise is no longer a slide. SAP is putting Claude inside the systems companies already use to close books, answer HR questions and reroute supplier orders.
The useful part of AI is not making more posts. It is changing how the work gets done.
AI becomes useful when it changes the operating system, not when it just produces more output.
Before adding volume, the site needed a clearer split between short notes, deeper insights and project workstreams.
A public thinking system needs different lanes for signal, synthesis and build logs.
Policies matter, but most teams do not yet understand the work they are trying to govern.
Before buying another tool or writing another policy, map the work.