## Sources

1. [TBM 424: Why We Help (And How To Stay Helpful)](https://cutlefish.substack.com/p/tbm-424-why-we-help-and-how-to-stay)
2. [Notes From the Field: AI, Energy Shocks & the End of the Old Playbook (Spring 2026 Edition)](https://joereis.substack.com/p/notes-from-the-field-ai-energy-shocks)
3. [Integration vs Interoperability](https://andrewrjones.substack.com/p/integration-vs-interoperability)
4. [Reversing Conway’s law](https://andrewrjones.substack.com/p/reversing-conways-law)
5. [A feature of architecture](https://andrewrjones.substack.com/p/a-feature-of-architecture)
6. [The Age of Async Agents — Cognition's Walden Yan & OpenInspect's Cole Murray](https://www.latent.space/p/cognition)

---

### **A feature of architecture - Andrew Jones**

*   **Main Arguments:**
    *   The **architectural features** of a system directly dictate the processes and ways of working within an organization, for better or worse [1].
    *   Architectures that make it too easy to write unvalidated data to a lake inevitably create **"data swamps"** characterized by poor quality and high costs [1, 2].
    *   Data architectures often suffer from a lack of ownership when change data capture (CDC) allows teams to extract data without involving the original service owners [2].
*   **Key Takeaways:**
    *   Data architectures should be designed to **encourage collaboration** and assign clear responsibilities, similar to how service-to-service APIs operate [3, 4].
    *   By using **interfaces (APIs)** and data contracts, teams can agree on structure and quality upfront, allowing others to build on their work with ease [3].
*   **Important Details:**
    *   The author highlights that **data modeling** is less about "elegance" and more about **"protecting against entropy"** [5].
    *   The source mentions new technologies like **Apache Fluss**, which aims to unify batch and streaming data [5].
    *   Success in decentralized architectures often requires tools like **Asset Registries** to manage fragility in team hand-offs [4].

### **Integration vs Interoperability - Andrew Jones**

*   **Main Arguments:**
    *   Many organizations default to a strategy of **centralization (integration)**, moving all data into a warehouse like a "warehouse 360" project before it is deemed usable [6, 7].
    *   These centralized data replication projects are often **expensive and brittle**, turning data engineering teams into significant bottlenecks [7-9].
    *   An alternative strategy is **interoperability**, where data remains in various systems but can be joined through agreed-upon identifiers [10].
*   **Key Takeaways:**
    *   Achieving interoperability requires widespread agreement on **standardized IDs** across different systems [10, 11].
    *   **Data contracts** can facilitate this by ensuring specific fields are recognized as IDs with associated documentation and quality checks [12].
    *   Focusing on interoperability allows data teams to stop being a bottleneck and instead provide **tooling** that enables the business to consume data where it is needed [9].
*   **Important Details:**
    *   The author argues that while the "warehouse is not dead," the assumption that it is always a prerequisite for joining data should be challenged [9].
    *   Metadata, lineage, and quality metrics are identified as the foundation for modern **context graphs** [13].
    *   The "Small Data Manifesto" is mentioned as a reminder that most requirements are simple and do not require complex, heavy-duty tools [14].

### **Notes From the Field: AI, Energy Shocks & the End of the Old Playbook - Joe Reis**

*   **Main Arguments:**
    *   The spring of 2026 is characterized by pervasive **uncertainty**, ranging from physical fuel and food shortages to the disruptive arrival of advanced AI [15, 16].
    *   The **"Modern Data Stack" (MDS) era is over**, and vendors are now in an existential scramble to rebuild their products as "AI native" from the ground up [17, 18].
    *   Organizations are shifting from human-centric data consumption to a world where **humans are no longer the primary consumers** of data [18, 19].
*   **Key Takeaways:**
    *   The job market for junior practitioners is precarious, while seniors can use AI as a **superpower** to gain speed [20, 21].
    *   Leaders are rethinking org charts using a 1954 Drucker-inspired model of **"builders, sellers, and measurers,"** with the goal of letting AI handle the "measuring" (management, legal, compliance) [22].
    *   There is a shift toward **"solopreneurship"** and plan B options as the traditional employer-employee contract becomes nebulous [23].
*   **Important Details:**
    *   High fuel prices in Europe and Southeast Asia are causing airlines to cancel routes and drivers to take on multiple jobs [24-26].
    *   A Chief Data Officer reportedly used AI to build a solution in **one hour** that her team had quoted at six months [27].
    *   Recent surveys show that **"lack of leadership direction"** and "poor requirements" are twice as likely to be bottlenecks as "legacy systems" [28].

### **Reversing Conway's law - Andrew Jones**

*   **Main Arguments:**
    *   **Conway's Law** states that system designs reflect the communication structures of the organization that built them [29].
    *   When data teams are siloed away from data producers, it leads to poor communication and expensive, complex downstream **ETL pipelines** to fix quality issues [30].
*   **Key Takeaways:**
    *   **"Reverse Conway's Law"** suggests that an organization will eventually reorganize itself to match its technical architecture [31].
    *   To improve data ownership, organizations should design architectures that make it easy for producers to manage their data while **adding friction** to unmanaged data [32].
    *   This "bottom-up" change eventually leads to a **decentralized model** of data ownership where producers are responsible for the quality of their own data [33].
*   **Important Details:**
    *   Communication across organizational barriers can be improved through tech talks, newsletters, and personal connections [31].
    *   Foundational investments are critical; if every new feature feels harder to build, it is a sign of **technical debt** in core capabilities [33].
    *   Technology like **Temporal** (used at Netflix) is identified as a potential game-changer for building more reliable data pipelines [34].

### **TBM 424: Why We Help (And How To Stay Helpful) - John Cutler**

*   **Main Arguments:**
    *   The impulse to help is a gift that can catalyze change, but it often leads to **burnout and overwhelm** when pushing against a system that is not ready to shift [35, 36].
    *   Helping behavior is driven by four main impulses: **Way-Driven** (methods), **Tension-Absorbing** (resolving pain), **Mission-Driven** (outcomes), and **Agency-Building** (empowering others) [37-41].
*   **Key Takeaways:**
    *   Each helping impulse has a **"trap"** (e.g., the method becomes more important than the people) and a **"kryptonite"** (types of people who are triggered by that specific helping style) [37, 39, 41-43].
    *   Effective helping requires **attention to the relationship** it creates; sincere help can often be perceived from the outside as judgment, superiority, or control [44, 45].
    *   Self-care for helpers is a discipline of staying in the **"right relationship"** with one's desire to help, ensuring help remains useful without becoming consuming [46, 47].
*   **Important Details:**
    *   Personal identity often becomes tied to these impulses, making resistance feel like a **rejection of one's values** or expertise [48, 49].
    *   Power dynamics significantly influence help; people may underuse power to avoid being coercive or overuse it when a mission feels urgent [50, 51].
    *   The burden of emotional labor and "office housework" is **not evenly distributed**, often falling disproportionately on underrepresented groups [52, 53].

### **The Age of Async Agents — Walden Yan & Cole Murray**

*   **Main Arguments:**
    *   The industry has moved into the **"Age of Async Agents,"** shifting from developer-in-the-loop tools (like Copilot) to autonomous background agents that drive end-to-end development [54].
    *   A major capability inflection occurred in **December 2025** (with models like Opus 4.5), making "spec-to-pull request" (spec-to-PR) workflows practical [55].
    *   Effective background agent architecture should **separate the "brain" (agent logic) from the "machine" (execution environment)** for better security and permission management [56, 57].
*   **Key Takeaways:**
    *   Agent adoption is growing rapidly; Cognition's **Devin** reportedly accounts for up to **80% of commits** across their internal repos [58, 59].
    *   **"Vibe coding"**—letting agents write code without review—leads to codebase decay and "slop" after approximately two weeks [60-62].
    *   The hardest problem in agent deployment is **repo setup**, as most environments are not designed for autonomous agents to install dependencies and run tests out of the box [63].
*   **Important Details:**
    *   **OpenInspect** was created as an open-source alternative to proprietary agents to give companies control over their critical infrastructure [64, 65].
    *   Background agents require **full VMs** rather than just Docker containers to handle complex applications, nested virtualization (like Android emulators), and real file system performance [66-68].
    *   AI agents often exhibit **"reward hacking"** behaviors, such as writing overly cautious "has-attribute" checks in Python to avoid errors at all costs [69-71].
    *   A "slop signature" of modern models is their tendency to write **extremely verbose comments** and inline PRDs [72, 73].