Daily Digest — 2026-10-26

The central premise of this video is that AI introduces a fundamentally new economic input, which the author calls intelligence, to enterprise operations. Traditionally, companies have managed two primary economic inputs: people and capital. While tokens are a measure of compute and seats are a SaaS measure for user access, AI itself represents intelligence, and its value should be viewed through a new economic lens.

The video highlights a common pitfall: enterprises getting stuck on traditional pricing and valuation models (e.g., per seat, per token) for AI solutions. Procurement teams, in particular, struggle with how to pay for AI, often reverting to familiar SaaS metrics. The author argues this is asking the wrong questions, as AI's true value lies in the intelligence it provides and the economic value it creates directly for the customer.

An example is given of a healthcare payments company, Cedar, using an AI agent named Kora to handle patient calls, check identities, explain charges, confirm insurance, and facilitate payments, even at 2 AM. The hospital, as Cedar’s client, doesn't care about the number of tokens Kora used; it cares that the bill got paid. This illustrates that the value delivered is the outcome, not the underlying compute or user interaction.

For CFOs, this shift in valuing AI presents a significant challenge. While companies may know what they're paying for in tokens, they often lack the know-how to earn a return on that investment because they don't understand how to allocate intelligence effectively. This changes the perceived value of a partnership.

For individuals within an organization, this creates a massive opportunity: to become a manager of intelligence. Instead of merely using AI to finish reports faster, individuals can leverage AI to find the root cause of customer problems and fix them. This requires a mindset shift from being a user of AI to being a manager of AI, deciding where and how the company's intelligence is deployed for maximum impact. This is something that everyone, from leaders to individual contributors, is figuring out together. Learning to turn AI resources into valuable work can significantly expand one's scope of responsibility.

The video also addresses why pricing comes up. The fundamental issue with purchasing intelligence as a business is its difficulty to value. Familiar pricing boxes like 'per token' or 'per seat' help procurement manage spend, but they don't guarantee useful thinking or outcomes. The burden of capturing enough value to justify the cost typically falls on the buyer. Many companies, in an effort to cap the spend and experiment cautiously, are under-invested in AI because they haven't figured out how to translate AI capabilities into ROI at scale. They lack the missing know-how to assign AI tasks effectively, change surrounding processes, or even evaluate the quality of AI-generated answers. A compelling AI demo doesn't equate to organizational readiness to use it effectively.

The concept of forward-deployed engineering is introduced as a pathway to bridge this gap. These are engineers who work closely with customers' organizations to turn technology into working systems. Their role is to help companies take purchased intelligence and turn it into value. OpenAI, for example, is described as a 'deployment company' with engineers working alongside business leaders and frontline teams to identify opportunities and build/test/deploy systems connected to customer data and processes. Anthropic also engages in similar forward-deployed engineering efforts through joint ventures with firms like Blackstone.

This leads to the idea of deeper partnerships, where commitment extends beyond mere usage metrics to focus on shared outcomes. While per-seat and per-token models exist, there's a growing trend towards outcome-based pricing or even joint ventures in cases where there's appetite for deeper commitment. Outcome pricing means a company pays when the AI delivers a result (e.g., a bill paid, a ticket resolved).

Such outcome-based models fit well in closed-loop scenarios (like the Cedar/Kora example). However, for more complex programs (like the equipment maintenance example where an engineer needs to distinguish between two seemingly similar faults), pure outcome pricing is harder to implement. Here, the focus shifts to ensuring that people are equipped to deliver value with the AI.

The author points to examples like Shopify and Block (formerly Square), which are highly AI fluent organizations. Shopify's COO Jess Hertz describes achieving a healthy return on AI spending, maintaining flat headcount while growing revenue by 34% (Q2 2026 vs. Q2 2025). Block has also cut its workforce and put AI at the heart of its business, showing 25% gross profit growth (Q2 2026). These companies demonstrate that high AI fluency enables buying usage and keeping the gains. For companies with low AI fluency, deeper partnerships, forward-deployed engineering, and integrated software + consulting are more suitable.

Ultimately, the call to action for individuals is to be a manager of AI – to lean into this role, decide what intelligence belongs where, and drive impact. The author encourages leadership to implement a revenue-project test: free up productivity, staff a new project that couldn't be resourced before, and attribute the resulting revenue to AI. This concrete thinking is crucial for moving beyond just cutting costs and instead leveraging AI for significant upside, as there's no ceiling on upside when innovation is embraced boldly.

The core of this video revolves around a remarkable demonstration of Agentic Engineering by OpenAI's 'AI Core and Experimentation' team. They tasked an AI agent with a broad, vague prompt: 'find unused code and clean it up.' The results were astonishing: over 1,400 pull requests were generated, leading to the deletion of 800,000 lines of code across various systems within approximately two weeks. This monumental effort was accomplished with the oversight of only five part-time engineers, showcasing a significant shift in how AI can contribute to software development.

Crucially, the AI agent's approach to identifying and deleting code went far beyond simple static analysis. It demonstrated advanced reasoning capabilities by checking a wide array of contextual information. This included all pull requests, Slack conversations, project trackers, feature flags, and even production profiling data. This comprehensive data analysis allowed the agent to accurately determine whether specific code sections were truly 'safe to remove' without adverse effects. The ability to integrate and synthesize information from such diverse sources is a hallmark of sophisticated agentic behavior, contrasting sharply with simpler code generation tasks often associated with AI.

The impact of this code cleanup was not merely cosmetic. The video provides concrete examples of the benefits. One container image size was reduced dramatically from 3 gigabytes to 1.5 gigabytes. Another notable achievement was the reduction of CI/CD time for a specific task from approximately 10 minutes to just 4 minutes. These are tangible improvements in infrastructure efficiency and development velocity that directly translate to cost savings and faster iteration cycles.

An interesting model economics observation was also highlighted. When comparing two different GPT models, GPT 6.1 Sol Ultra found 130 issues, while GPT 6 Astra Ultra found 140 issues. However, the GPT 6.1 Sol Ultra operated at about 14% of the API cost of GPT 6 Astra Ultra, effectively making it 10 times cheaper. This illustrates a key lesson in Agentic Engineering: it's not always about using the 'smartest' or most expensive model. Instead, the strategy should be to 'find the cheapest model that reliably clears your review threshold, learn from where it fails, improve the instructions, and then scale.' This iterative refinement of agent instructions and model selection is critical for optimizing both performance and cost.

The video emphasizes that humans remain 'very much in the loop.' The AI agent's role is not to operate autonomously without oversight. In one instance, the agent identified a 'supposedly permanent feature flag' that was wasting CPU resources. Instead of blindly removing it, the agent surfaced the ambiguity for review to the engineer. This human-in-the-loop interaction is vital for handling complex decisions and ensuring the integrity of the codebase, demonstrating that AI agents are powerful tools for augmentation rather than complete replacement.

This marks a significant shift in Agentic Engineering. The job isn't just about getting an AI to write new code; it's about 'giving an agent enough context, evidence, and boundaries so that it can investigate what should change in the first place.' The article author, Rodrigo Romes, is credited for his detailed work, including the actual prompts used, which further democratizes the understanding and application of these techniques. The ultimate question posed is whether coding agents will eventually create more software than they maintain, or if maintaining and simplifying existing, often complex, human-written systems will become one of their most valuable jobs in the future.

The presentation, "How to Build Financial Agents that Work Across the Web," by Jay Sahnan and Arik Bird from Browserbase, focuses on overcoming the challenges of automating tasks on the internet for financial institutions.

### What is a Browser Agent? A browser agent is an AI agent that uses a browser to read pages, choose its next step, and take action. It's essential for tasks that involve interacting with websites that lack APIs or have complex user interfaces, such as booking flights, planning events, or gathering competitor pricing. Many financial tasks fall into this category, requiring access to 'hardest to reach' parts of the web, often behind login portals or protected by anti-bot measures.

### Challenges of Web Automation for Financial Institutions The speakers highlighted three prominent financial companies – Ramp, Bill, and Stripe – as examples of organizations facing significant challenges in automating web-based financial processes. For Ramp, the core problem was collecting missing receipts from various merchant portals. Initially, they resorted to manual human teams, but this approach proved unsustainable and unscalable as the company grew. The solution lay in automating this process reliably at scale.

### Understanding Chromium and Web Fingerprinting To understand how to automate reliably, the speakers delved into how Chromium (the open-source base for Chrome and other browsers) operates. When a human accesses a website, their browser presents a consistent fingerprint, including browser version, OS, fonts, screen size, and timezone. The website also recognizes a familiar visitor through cookies, login state, and previous visit history. Finally, there's a familiar pattern in network reputation, request timing, and navigation, all indicating human interaction.

### The Problem with Traditional Headless Browsers in Data Centers When trying to scale automation by running a headless Chromium binary in a data center, these identifying factors change significantly: Different fingerprint: The browser's characteristics will vary. Fresh cookies and login state: Each session starts clean, requiring repeated logins. * Data center IP: The network originating from a data center is a clear indicator to anti-bot systems that it's not a human residential IP. While proxies can mask the IP, they don't solve the underlying fingerprinting or behavioral differences.

The fundamental problem is that using the same Chromium binary designed for a laptop in a data center exposes it as non-human due to these environmental differences. Furthermore, Chromium undergoes changes every two weeks (new releases) and weekly security updates, making it a moving target for automation.

### Browserbase's Solution: Forked Chromium and Stagehand Browserbase addressed these issues by forking Chromium and making three key changes: 1. Remove the automation flag: They modified the source code to change the `navigator.webdriver` getter to always return `false`, preventing websites from easily detecting automated browsing. 2. Align browser identity: They made numerous internal changes to ensure the browser identifies itself as a standard Chrome browser, not a headless one. This involves aligning headers, web APIs, and other client-side properties. 3. Keep the process alive: Instead of launching and killing a Chromium instance for every task, Browserbase maintains persistent browser processes and closes individual windows, improving efficiency and mimicking human browser usage patterns.

### Scaling and Controlling Browsers with Stagehand With these modifications, Browserbase can launch thousands of verified browsers concurrently that behave indistinguishably from human-operated ones. For Ramp, this enabled automated, reliable receipt collection, saving over 30,000 manual data entry hours and achieving a 5x increase in scale.

For Bill, the challenge was not just scale but control over diverse vendor portals, each with a different payment flow (e.g., varying fields, steps, and UI). Traditional Playwright scripts, which rely on hard-coded selectors, break easily when a page changes. Browserbase's Stagehand SDK (an open-source solution built on top of Playwright) offers an improved approach: Natural Language Instructions: Instead of hard-coded selectors, agents can receive natural language prompts (e.g., "Find the learn more button," "Click the login button," "Extract product details"). Smart DOM Processing: Stagehand reads the page structure (DOM and accessibility tree), identifies relevant elements, and preserves targets for actions. Critically, it shrinks the DOM context sent to the LLM from potentially 122,000+ tokens (raw HTML) down to ~4,000 tokens for the same example. This significantly reduces LLM costs and improves response times. * Fewer Network Trips: Stagehand runs as a Chrome extension directly alongside the hosted browser. This co-location reduces network latency between the agent's instructions and browser actions, leading to 1.59x higher throughput compared to traditional remote browser control (e.g., Playwright + Selenium).

### Advanced Use Cases with Stripe For Stripe, the challenge involved dynamic decision-making based on research findings. Stripe needs to onboard tens of thousands of merchants, requiring verification of business legitimacy through various sources like company websites, investor relations pages, SEC filings, and stock exchange listings. An agent's next step (e.g., checking a policy page, a review site) depends on what it finds. Stagehand's `observe`, `act`, and `extract` primitives allow for this dynamic behavior: `observe`: Finds the next place to look (e.g., `await stagehand.observe("Find the learn more button")`). `act`: Executes an action (e.g., `await stagehand.act("click the login button")`). * `extract`: Turns page content into structured data using a zod schema (e.g., extracting product name, price, and stock status). This is highly valuable for engineers, converting messy unstructured web data into usable formats.

The demonstration showed Stagehand concurrently researching four different sources for a company (website, investor page, SEC filing, NYSE listing) and then scaling this to 20 companies, utilizing 60 browsers. The live demo highlighted the efficiency and reliability of collecting structured data from diverse and complex web sources, often faster than the real-time stream could visually update. This robust, unblocked automation is made possible by Browserbase's verified browser environment and Stagehand's intelligent interaction capabilities.

This video demonstrates how Superblocks enables efficient and secure development of AI-powered internal applications, particularly focusing on the needs of a large enterprise. The core problem addressed is the current disconnect between departmental teams building AI-powered apps with various tools and IT's lack of visibility and control over these applications, leading to potential security and compliance risks.

The Customer Support CRM Example: The video showcases building a Customer Support CRM using Superblocks. This CRM manages customer details, support tickets, and refund requests in one centralized platform. The key functionalities highlighted are:

Customer Search and Management: Easy access to customer information. Ticket Management: Opening and tracking support tickets. * Refund Processing: Initiating and approving refund requests.

Addressing Role-Based Permissions and Workflows: The video emphasizes the importance of role-based permissions in enterprise applications. It illustrates how different roles (e.g., support agent vs. manager) have distinct access levels and functionalities within the CRM. For instance, a support agent can request a refund, but only a manager can approve it. This ensures that business rules are enforced and audited.

Superblocks' AI Assistant – Clark: Superblocks integrates an AI assistant named Clark to streamline the development process. Clark aids in planning the application by interpreting natural language descriptions of the desired features and workflows. It also pulls information from the company's knowledge base, such as a design system, to ensure consistency in the application's look and feel with other internal tools.

Automated Security and Compliance: A critical aspect highlighted is Superblocks' focus on security and compliance, which is often overlooked in rapid AI app development. Key features include:

Automated Security Checks: Superblocks automatically runs security scans at multiple stages (build, publish, and in production) using its Security Agent and Vulnerability Detection tools. This proactive approach identifies issues like trusting client-side identity information or insufficient validation of refund amounts against original order totals. Policy Agents: These agents enforce predefined security policies on the generated code, ensuring that sensitive actions (like approving refunds) require server-side authorization and proper verification of user roles and data integrity. Audit Logs: The platform provides comprehensive audit logs that record every change and action within the application, including who performed the action, the resource affected, and the outcome. This offers IT visibility into how the platform and applications are being used and modified. Security Center: A centralized hub where IT administrators can monitor all applications, track security runs, identify vulnerabilities, and manage exceptions. This provides air traffic control for internal apps, allowing IT to oversee who builds, ships, and changes applications across the organization.

Cloud-Prem Deployment Model: Superblocks offers a Cloud-Prem deployment model, where all Superblocks services are hosted within the customer's private AWS VPC. This design ensures that all application data, AI inputs, and AI outputs remain within the customer's control and environment. AI inference is executed on Amazon Bedrock, utilizing models and regions controlled by the customer. This addresses major concerns regarding data residency, privacy, and intellectual property protection for enterprises.

Key Takeaways for IT and Development Teams: Empower Business Users: Superblocks allows business teams to build useful AI applications quickly without deep technical expertise. Maintain IT Control: IT teams gain full visibility and control over internal AI apps, including access management (SSO login by default), security, and deployment. Secure Development Lifecycle: Automated security checks and policy enforcement are integrated into the development and deployment process, reducing risks and ensuring compliance. Data Sovereignty: The Cloud-Prem model ensures sensitive company data remains within the customer's private cloud, preventing data exfiltration and maintaining compliance with industry regulations.

The video concludes by emphasizing that Superblocks provides a crucial third option beyond blocking AI app development or allowing uncontrolled shadow IT. It allows both business teams and IT to collaborate effectively, fostering innovation while protecting data and ensuring security.