Newsletter #128
Enjoy Newsletter #128.
Some new agentic patterns
Jesse Vincent, creator of the open-source agent SDK Lace, shares the new patterns his company Prime Radiant uses to integrate AI agents into daily workflows. Rather than waiting to be prompted, these agents work proactively right inside Slack: Scribble monitors emerging issues and autonomously updates internal documentation, Nora supports go-to-market efforts and suggests opportunities like conferences worth submitting talks to, and Sen is a suite of personal assistant agents handling email triage, scheduling, and research. The standout pattern is what the author calls “agentic user in the loop”: while developing Sen 2.0, Claude Code worked directly with another agent named Ada-sen without a human relaying messages between them. Claude proposed changes, Ada read the spec, asked clarifying questions, ran live tests, and gave feedback, with the two iterating back and forth—sometimes shipping entire features overnight. For security, the architecture relies on privilege separation: the primary agent cannot make arbitrary outbound connections, credentials live in 1Password and are injected into requests only through an authenticating proxy, and any subagent that needs them must request approval from a “judge” agent, keeping secrets out of conversation transcripts. Even so, Vincent acknowledges that Simon Willison’s “Lethal Trifecta”—an agent that can read private data, communicate externally, and ingest untrusted content—remains an ever-present risk requiring careful mitigation.
Why I Stopped Arguing With People
Software engineer Cong Wang reflects on his evolution from loving technical debates to prove he was right, to realizing that most arguments achieve nothing. Drawing from Laozi’s Tao Te Ching—where “being” and “non-being” give rise to each other—he notes that for someone to be right, someone else must be cast as clearly wrong. Winning an argument creates a loser, making the victory hollow. In his view, most arguments defend egos rather than ideas: once the other person feels personally attacked, they dig in, and the sharper your logic, the harder their position calcifies.
Wang points out that humans feel first and rationalize second, leaving pure logic powerless against emotionally rooted beliefs. People rarely learn from unsolicited advice; they learn from consequences, and correcting others usually breeds resentment. The only exception is when someone actively asks for help—only then do their defenses drop so real learning can happen. Instead of trying to win debates, he suggests treating disagreement as a competitive advantage, since business opportunities often lie precisely where consensus hasn’t formed yet. Ultimately, the only person you can change is yourself, by humbly seeking feedback and genuinely taking it to heart.
Understanding is the new bottleneck
In his July 2026 AI Engineer talk, Notion design engineer Geoffrey Litt argues that as AI agents generate more and more code, the primary bottleneck is no longer typing speed—it is human understanding. According to Litt, understanding the codebase isn’t mainly about catching agent mistakes (which agents are increasingly good at verifying themselves), but about maintaining enough mental fluency in the system to keep contributing good ideas across dozens of iterative rounds with the agent. Skipping comprehension accumulates “cognitive debt” that compounds painfully over time.
Litt highlights three techniques to address this. First, rich code walkthroughs: his custom /explain-diff skill generates explanations that lead with background context, intuition before low-level details, and interactive diagrams, paired with comprehension quiz questions that act as a “speed bump” forcing the author to truly understand the change before sharing it. Second, “micro-worlds,” inspired by educator Seymour Papert’s idea of “living in Mathland”: having agents spin up small custom interactive environments—like visual debuggers, dashboards, or simulations—so humans can explore and build intuition about how the system behaves. Third, shared workspaces where agents plan alongside humans on collaborative docs so the whole team thinks together rather than in isolated silos. He closes with Alan Kay’s 50-year-old vision: computers exist to augment human intellect, not merely to automate tasks away.
When to repeat yourself
Franco Fernando revisits the DRY (Don’t Repeat Yourself) principle, arguing that deduplicating code carries hidden costs many engineers overlook. The biggest is the “coordination tax”: when multiple teams share a module, every change requires alignment meetings, design reviews, and merge negotiations, whereas a team modifying its own local copy can move immediately. The real question, then, is whether coordinating a shared abstraction costs more than letting separate copies drift apart. Fernando distinguishes incidental duplication—code that looks similar today but belongs to different business domains and will evolve in different directions—from essential duplication, where the exact same business rule is repeated in multiple places; only the latter belongs in a shared abstraction.
He advises patience, because waiting for clarity is cheap while premature abstraction is expensive. Every code-sharing mechanism comes with trade-offs: shared libraries are easy to extract but create transitive dependency and versioning headaches; microservices offer clean API contracts but add network latency, operational overhead, and availability risks; inheritance breeds rigid hierarchies, while composition is more flexible at the cost of extra wiring. His rule of thumb: within a single team’s boundary, DRY usually wins, but across organizational boundaries you should recalculate the trade-off and favor reversible choices—because merging duplicated code later is far easier than untangling a tightly coupled abstraction.
Let AI Burn
Ed Zitron argues that the AI industry is a bubble built on hype and circular financing, and if it bursts, the government should not bail it out. In his view, the situation bears no resemblance to 2008: banks were vital economic infrastructure, whereas AI companies are not, and aside from SoftBank lending over $40 billion to OpenAI, there is no comparable systemic contagion risk. In 2026 alone, OpenAI and Anthropic raised over $300 billion while continuing to post massive losses; those two companies account for 89% of revenue among top AI startups, while the rest of the sector generates roughly $20 billion a year combined. Zitron also rejects the familiar dot-com bubble analogy: overbuilt fiber-optic cables remained useful for decades, whereas GPUs are narrow-purpose assets that depreciate rapidly. Hyperscalers are projected to spend $765 billion in capex in 2026 and over $1 trillion in 2027, with OpenAI alone burning roughly $50 billion on compute this year. He dismisses Sam Altman’s proposal to give the US government a 5% equity stake worth $42 billion, noting that figure wouldn’t even cover a single year of compute spend and does nothing to fix underlying unprofitability. His conclusion: let the market clear itself, because any bailout would require endless capital injections with no exit in sight.
When AI Costs More Than the Engineer
Tomasz Tunguz points out a striking inversion in software economics: at leading AI-native companies, compute spend per employee has surpassed compensation. Anthropic spends roughly $2 million per employee annually on compute—2.3 times its fully burdened compensation cost of around $500,000 per head. In return, revenue per employee reaches $14 million at Anthropic and $6.5 million at OpenAI, the highest in the world and far above the $250,000–$600,000 typical of traditional SaaS companies.
The rest of the market is still far behind. The top 1% of software companies spend roughly $89,000 per engineer per year on AI—about 40% of a $224,000 salary—while the median company spends just $137. Tunguz models three scenarios for 2029: a low case of $106,000 per engineer (41% of salary), a base case of $363,000 (140%), and a high case of $596,000 (230%), matching Anthropic’s ratio today. Where the market lands depends on two opposing forces. On one side, token prices have fallen roughly 10x per year for the past three years, open-weight models are narrowing the quality gap at lower cost, and enterprises can cap usage by role. On the other side, frontier model pricing holds steady while agentic workflows consume vastly more tokens—to the point that Goldman Sachs projects token consumption will grow 24x by 2030.
The best code is the one you shift+delete
RavenDB creator Oren Eini argues that measuring coding models by how fast they write production code misses the point; their real superpower is making previously impractical, throwaway work effortless. He recounts debugging a production incident involving 25–30 MB of compressed logs. Finding individual errors was easy, but correlating events across a timeline was hard, and feeding the entire log into an LLM was both impractical and expensive. Instead, he pasted the first ten lines and asked the model to write a script to extract, aggregate, and tabulate the events; in under a minute he had a tool he could iterate on, such as plotting index changes over time. The root cause jumped out immediately: the customer was running multiple application versions, each with its own index definitions, continuously overwriting one another. Eini never read a line of the script and deleted it as soon as the bug was found.
His second example is the polar opposite: refactoring query execution in Corax, core engine code meant to live in the product for decades. Here the model typed while Eini steered and reviewed every line. He warns that after hours of trying, reverting, and retrying approaches, it is all too easy to look up and find two thousand lines of diff you don’t truly own. To retain authorship and confidence, he relies on existing regression tests for old behavior, new tests for new behavior, and a disposable test harness that runs the old and new implementations side-by-side and visualizes the results. Building temporary scaffolding like that used to take hours or weeks; now it takes minutes.
Bonus
Articles: Chia sẻ suy nghĩ về sản phẩm và người làm sản phẩm
Vuong Quang Khai, founder of Zalo, shares his perspective on product building: the three universal product values are usefulness, aesthetics, and intuitiveness, to which Zalo adds simplicity, reliability, and privacy—and the hardest part of the craft is making trade-offs when good values conflict. In product builders, he prizes three qualities above all: deep user empathy, the courage to make tough trade-offs, and obsession with small details, while expressing skepticism toward purely data-driven decision-making or A/B testing and emphasizing that, above all, you must genuinely love the product you build.
This post was rewritten by Claude Code with Opus 5.5 on 27/09/2026.
Comments