Quanta Bits - April 18, 2026

Hi everyone,

Busy week, mostly heads down in work, catching news in headlines. Opus 4.7 dropped. Claude Code kept packing in features. The kind of week where the big stuff sneaks past you unless you look up.

Then, logging off Friday evening, Salesforce announced something we all have been suspecting was happening already: they are making the entire platform accessible through agents and APIs, no browser required. A Tier-1 SaaS leader admitting, out loud, that the interface is not the moat anymore. The trend had been obvious. Seeing Salesforce commit to it in public was still shocking.

That is this week's essay.

1. The AI layoff thesis went from one-off to pattern. Snap announced it was cutting 1,000 roles, roughly 16% of the company, explicitly tying the decision to AI now generating 65% of its new code. Stock rose 7 to 9% on the announcement. Same playbook Jack Dorsey ran at Block in February: 40% cut, 20%+ stock pop, same public AI-substitution framing. Two data points is not yet a trend, but the market reaction is the tell. Public equities reward CEOs who say the AI-substitution part out loud. Expect more boards to pressure their management teams to name the specific headcount AI is supposed to absorb, on the record (Tech Startups, April 15).

2. CFOs are doubling AI spend while most cannot show ROI yet. Silicon Valley Bank's April survey of 230 VC-backed CFOs shows median AI spend jumping from $20,000 to $50,000 per month. Grant Thornton's 2026 AI Impact Survey this week found 78% of executives lack confidence they could pass an independent AI governance audit within 90 days. The kicker from Grant Thornton: programs owned by the CFO are 1.4x more likely to report significant value than those owned by the CIO. Finance is inheriting AI accountability faster than the governance plumbing can keep up. If your CFO is not an AI program co-owner yet, that is this quarter's conversation (CFO Brew, Grant Thornton).

3. Agent identity is becoming its own enterprise security category. Okta published research on April 13 arguing AI agents are now acting like employees but organizations still treat them like software. The same week, SailPoint was named Gartner Peer Insights Customers' Choice specifically for Agent Identity Security and Shadow AI Remediation. Gartner projects 40% of enterprise applications will embed task-specific agents by end of 2026, up from less than 5% in 2025. IAM and PAM budgets are now competing directly with model spend. If your agents are not in your identity directory with owners and revocation paths, they are shadow employees with production access (Fortune / Okta, SailPoint via StockTitan).

4. The executive-manager AI perception gap got quantified. Harvard Business Review published a Wharton / GBK Collective study of U.S. companies this week showing a consistent split: 45% of executives report significantly positive AI ROI, only 27% of middle managers do. 56% of executives say they are moving faster than competitors, only 28% of managers agree. Companion BCG / Columbia data: 76% of executives believe employees are enthusiastic about AI; only 31% of individual contributors actually are. This tracks everything I have seen in the field. Managers are not the blocker. They are the accurate signal, deploying AI into real workflows with uneven technical comfort and output that has to be consistently right. If your exec deck says the rollout is going well and your managers say it is not, believe the managers (HBR).

On Friday at its TDX developer conference in San Francisco, Salesforce announced Headless 360, which VentureBeat called the biggest architectural change the company has made in 27 years. Every capability in the platform is now available as an API, MCP tool, or CLI command. Developers can build on Salesforce from Claude Code, Cursor, Codex, or Windsurf without opening a Salesforce browser. A new Experience Layer renders agent output natively inside Slack, Teams, ChatGPT, Claude, Gemini, and mobile. The official Salesforce framing: "Instead of burying capabilities behind a UI, expose them so the entire platform will be programmable and accessible from anywhere." One early customer, a travel company called Engine, built an Agentforce agent in twelve days that now handles half of its 800,000 annual chat cases and is expected to save $2M a year (Salesforce Ben).

Salesforce is the first Tier-1 SaaS vendor to make this its defining architectural bet. Others have been moving in the same direction quietly for months. The UI was never the thing Salesforce won on. It won on data model, integrations, the AppExchange, and the fact that it was a real cloud platform when everyone else was still selling on-prem CRM. But the UI was the front door, what users opened every morning, what made adoption either work or stall. More importantly, it was how SaaS vendors stayed in the daily relationship with their customers.

To use Helm's Deep analogy, Headless 360 is Salesforce retreating from the outer wall to defend the keep. The UI was the first line, the place every interface company fought for. Salesforce just decided that line is no longer worth holding. Salesforce kept what sits behind the keep: the customer data, the relationship graph, the sales motions, the institutional memory. Someone else gets the daily interaction. In January, Marc Benioff told the New York Times DealBook that the enterprise AI stack has three layers: a data layer that harmonizes and integrates, an app layer, and an agentic platform on top. Headless 360 is that framework shipped as product. It is also an admission: if the UI becomes a commodity, Salesforce intends to be the data and governance layer that every agent has to query.

The peers have been moving the same way, more quietly. HubSpot shipped a production MCP server with full write access to its CRM (HubSpot). SAP is rolling out MCP across HANA Cloud, ABAP, and Commerce Cloud. Microsoft's April Copilot wave adds MCP for declarative agents and agent-to-agent communication (Microsoft). Anthropic has been packing enterprise workflows into Claude directly with Cowork plugins for Google Workspace, Docusign, FactSet, Harvey, and private marketplaces for regulated firms (TechCrunch). Satya Nadella said the quiet part out loud on Microsoft's most recent earnings call: "The corporate moat has a new definition, from owning data to owning a business-savvy model." The interface layer is being absorbed into frontier AI tools. Every SaaS vendor is now making the same decision Salesforce just made: are you a UI or a data layer?

Headless 360 exposes APIs, not a chat mandate. Dashboards and custom interfaces can consume the same endpoints. But Salesforce chose to prioritize rendering into Slack, Teams, ChatGPT, Claude, and Gemini. That is a chat-forward bet, and chat is not always the right interface. Scanning a pipeline view, spotting the outlier deal in a list of fifty, comparing two quarters side by side, these are visual tasks that a dashboard handles better than a conversation. A 2025 METR study found experienced developers were 19% slower using AI chat interfaces than they predicted. A 2026 follow-up showed the same developers now 18% faster. The chat paradigm is maturing, but the jury is out on where the line sits.

Enterprise contracts also buy time. Most SaaS agreements run three to five years, which means even if the direction of travel is right, incumbents collect the revenue all the way through the transition.

Taken seriously, those counters sharpen the thesis rather than break it. My weight is that this is a trend, not a blip. I have spent the last twenty years watching sales, marketing, and customer success teams fight their own software. They do not love Salesforce. They use it because they have to. And the thing chat does better than any dashboard is consolidate. When you are prepping for a customer meeting, the answer lives in Salesforce opportunities, Service Cloud tickets, email threads, a Slack channel, a Notion page, and a Gong call. A chat interface pulls all of that into one coherent answer. A dashboard shows you one slice at a time. The right mental model is not agent replacing software. It is a conversation with an analyst who does the work after. That is what Slack plus Agentforce is selling. That is what Claude Cowork is selling. At some point, mid-market buyers stop caring where the customer history physically lives, as long as the answers they get are right. Not in regulated workloads. Not in the next twelve months. But inside the window of every software decision being made this year.

Salesforce knows this. You do not unbundle your own product unless you see the threat. The bet here is an Intel Inside play: invisible but essential. The losers will not be Salesforce, SAP, or Oracle. The losers are mid-tier horizontal SaaS, the Zendesks and Asanas and Monday.coms and Gainsights, vendors without a vertical specialty or a unique data asset. Public markets have not priced this split yet. 2026 has been brutal across SaaS: the broad software ETF is down more than 20%, and vertical and horizontal names have both been hit. The market is still treating SaaS as one category under AI pressure. I think that is wrong, and the early private-market data backs that up. Bessemer's January 2026 paper on vertical AI found LLM-native vertical companies growing 400% year over year at 65% gross margins (Bessemer). The surviving SaaS archetype is an enterprise infrastructure company with real depth in one vertical or one functional horizontal, defended by at least two of four things: proprietary data, embedded workflows, a regulatory barrier, or a network effect. Salesforce has all four. A generic horizontal productivity tool has one or zero.

Which means every enterprise software RFP from here on is really one question. Does this vendor own defensible data or a defensible workflow, or am I paying for a UI that Claude can replicate in six months? The UI was never the moat. It was the billing surface. Salesforce just admitted it out loud.

From wool runners to GPU runners. Allbirds, the sustainable sneaker company that went public at a $4.1B valuation in 2021, announced on April 15 that it is selling its entire footwear business and rebranding as "NewBird AI" to pivot into GPU compute infrastructure. Stock jumped roughly 600% on the news. If the market will pay a 6x premium for a company that no longer makes anything, vendor due diligence just got harder. Dot-com bubble anyone? (Bloomberg).

AI as systemic financial risk, on the record. On April 7, Treasury Secretary Scott Bessent and Fed Chair Jay Powell summoned the CEOs of Bank of America, Citigroup, Goldman Sachs, Morgan Stanley, and Wells Fargo to Treasury headquarters to discuss cybersecurity risks from Anthropic's Claude Mythos model, which has demonstrated an unprecedented ability to identify zero-day vulnerabilities. First time US regulators have convened bank chiefs over a single AI model release. Jamie Dimon, who could not attend, flagged AI "cybersecurity vulnerabilities" in his annual shareholder letter the same week. For risk leaders, "AI as systemic risk" just moved from think-piece to regulatory agenda (CNBC).

Anthropic is no longer a model company. Two shipments in two weeks prove it. Claude Managed Agents handles the sandboxing and tool orchestration work that used to justify "we need six more months before we deploy agents." Claude Design recreated a landing page from a one-line prompt and produced a 30-second animated launch video in four minutes. Between these, Claude Code, and Claude Cowork, Anthropic is now competing with Figma, Microsoft Office, Google Workspace, and Replit at once. The infrastructure excuse is gone. The remaining gap is organizational (SiliconAngle, Peter Yang).

Source: Superhuman AI Newsletter

▶️ Design Is the New Code, Dylan Field (Figma CEO) with Peter Yang. Field's framing: AI output is clay, not a final answer, and real craft is what you do with it from there. The bit about 60% of Figma designs now being made by non-designers landed. Worth an hour if you are thinking about how builders' roles are merging.

▶️ A2A vs MCP: AI Agent Communication Explained, IBM Technology. The clearest primer I have found on the two protocols defining how agents talk to each other. A2A lets agents from different vendors exchange tasks. MCP connects agents to tools and data. Not either/or, most real deployments need both. If your team is starting to ask "how do our agents work together," this is the 10-minute explainer to share.

🎙️ How to Build a Personal Context Portfolio and MCP Server, Nathaniel Whittemore (The AI Daily Brief). Ten structured markdown files as an operating manual for every AI you work with, portable across Claude, ChatGPT, Gemini. The MCP deployment piece is the practical punchline: your context as a server any agent can call. Worth reading before you explain to clients why their AI tools keep forgetting who they are.

James Islington — The Will of the Many (2023) | 639 pages | ★★★★★     The Strength of the Few (2025) | 725 pages | ★★☆☆☆

The Will of the Many grabs you on page one and doesn't loosen its grip. A Roman-inspired empire built on a magic system where people cede their life force up a pyramid, and a smart orphan with a hidden identity infiltrating its most elite academy. Relentless pacing, clean prose, layered political intrigue, and a protagonist who is a genuine pleasure to follow. Even the wolf pup earns his keep. Couldn't put it down.

The Strength of the Few is the disappointing follow-up. Book one's tight thread splinters into three parallel-world timelines that never quite cohere, with too many new characters and battles that blur together. A few returning threads land, but the three-world structure works against the story, dispersing momentum instead of building it. Finished it more out of obligation than enthusiasm. A promising series that lost its footing in the second step.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖