X Tutup
Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 24 additions & 0 deletions .scribe/beyondthecode-journal.md
Original file line number Diff line number Diff line change
Expand Up @@ -53,3 +53,27 @@
**Learning:** Initial hero image used made-up TypeScript about "feature velocity" and "comprehension metrics." Felt fake. Replaced with real Python — an async connection pool with semaphores and locks. The critical section (race condition handling) blurs out. Real code that engineers recognize is more effective than code that illustrates the essay's concepts literally.

**Implication:** Visual elements should ground the essay in recognizable reality, not mirror its abstractions. Show production code, not conceptual code.

---

## 2026-02-24 – The Authorization Bottleneck and Judgment Capital

**Learning:** "Productivity" metrics in AI-assisted environments are dangerously incomplete because they only track the collapse of production costs, while ignoring the fixed (or increasing) cost of authorization. Every unit of unverified AI output creates an "audit debt" that senior engineers must absorb. If the organization only rewards the generator, the auditor becomes a structural blocker, leading to a rational but dangerous compression of review depth.

**Implication:** Future essays should focus on the "sign-off" layer rather than the "typing" layer. The real shift is not in how much we can build, but in how much we can safely verify.

---

## 2026-02-24 – Vibe-Coding as Shadow IT Revival

**Learning:** AI has lowered the barrier to software construction so far that "Shadow IT" has returned as a team-level productivity hack. Bespoke tools (vibecoded dashboards, CRM bridges) are trading Organizational Legibility for Task Velocity. These tools are only "disposable" until they become critical path, at which point they become unmanaged maintenance liabilities that Staff engineers eventually inherit.

**Implication:** When evaluating "productivity gains" from AI internal tools, look for the latent maintenance cliff. The saved vendor cost is often just a deferred complexity tax.

---

## 2026-02-24 – Forced Documentation through Agent Context

**Learning:** A surprising inversion: because AI agents require high-fidelity state to function across sessions, developers are being forced back into disciplined documentation (ADRs, AGENTS.md, decision logs). This is not a cultural change but a functional necessity for tool performance. The agent's need for context is achieving what years of engineering management could not.

**Implication:** Monitor the shift from "slop" to "specification." The most successful teams might not be those with the best prompts, but those with the best context management.
73 changes: 73 additions & 0 deletions src/content/beyondthecode/the-authorization-gap.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,73 @@
---
title: "The Authorization Gap"
date: 2026-02-24
description: "How the decoupling of code generation and code judgment creates a new organizational bottleneck that velocity metrics cannot measure."
author: "Ganesh Pagade"
draft: false
---

<p class="drop-cap">The junior engineer submitted a three-thousand-line Pull Request on Tuesday morning. By Tuesday afternoon, he had submitted two more. The velocity dashboard for the squad showed a vertical spike. In the quarterly business review, the Director would point to this chart as evidence of a 10x productivity gain through AI leverage.</p>

On the other side of the screen, the Senior Engineer responsible for reviewing those changes stared at the notification queue. She had her own features to ship, three incident post-mortems to write, and a promotion calibration meeting in an hour.

**The cost of generating code has collapsed, but the cost of authorizing it has remained fixed.**

## The Sign-off Bottleneck

In the manual era, production and judgment were coupled by the speed of human typing. It took a day to write a few hundred lines of logic, and it took an hour to review them. The reviewer’s bandwidth was rarely the primary constraint because the author’s production speed provided a natural rate-limit.

AI-assisted development has decoupled these processes. An engineer can now generate complex implementations faster than a peer can critically audit them. The bottleneck has shifted from the "typing fingers" to the "authorizing eyes."

**This is the Authorization Gap: the widening distance between what an organization can produce and what it can safely verify.**

When an organization measures "productivity" solely by output, it incentivizes the behavior that compounds this gap. The junior engineer is rewarded for the volume of his submissions. The senior engineer, however, is not explicitly rewarded for the depth of her audit. In a system that stack-ranks by visible impact, the auditor becomes a "blocker" to the generator's velocity.

## The Liability of the Master Key

The common argument for AI leverage is that it "unblocks" the developer. But in high-stakes corporate environments, speed was rarely the primary constraint that justified the "locks" of code review, CI/CD gates, and staging environments. Those locks exist to manage the liability exposure of the employer.

Every line of code is a liability until it is proven to be an asset. **AI-assisted development provides a fire hose of potential liability that anyone can point at production.**

The organizational risk is not that the AI will write bad code—it is that the human judgment layer will fail to catch it under the pressure of inflated velocity expectations. If a Senior Engineer is expected to maintain her own output while reviewing 5x more generated code from her squad, the depth of her review often compresses.

She faces a rational but dangerous choice: trust the generator to avoid becoming a bottleneck, or maintain the standard and fall behind on her own metrics. Most choose a middle path of "vibe-checking"—scanning for patterns rather than verifying logic. This is where structural risk accumulates.

## The Calibration Distortion

This shift creates a profound distortion in performance calibration.

Traditional metrics like DORA or story points assume that output is a proxy for impact. This assumption held when output required a corresponding amount of human reasoning. But when output can be generated with minimal engagement, the proxy breaks.

A Junior Engineer who ships ten AI-assisted features might appear more "impactful" on paper than a Staff Engineer who ships two manually-reasoned architectural changes. In a calibration meeting, the Director sees the volume and the speed. They do not see the "audit debt" the Staff Engineer absorbed to ensure those ten features didn't break the payment gateway.

**The organization effectively taxes its most experienced judgment to subsidize its least experienced output.**

If the promotion system rewards the generator and ignores the auditor, the most rational move for a Senior Engineer is to stop auditing deeply and start generating more slop of her own. The organizational knowledge that traditionally formed through the friction of manual review begins to evaporate.

## The Illusion of Zero Cost

When a developer uses AI to "vibe-code" a custom internal tool instead of buying a vendor solution, the immediate balance sheet looks improved. The "build vs. buy" decision favors "build" because the labor cost appears to have vanished.

But authorization cost is rarely zero. A "vibe-coded" dashboard that manages production deployments is still a production system. It requires maintenance, security auditing, and long-term ownership.

The organization discovers the cost of the "cheap" code six months later, when the author has moved to a different team and the system fails during a 3 AM incident. The on-call engineer discovers they are debugging a black box written by a black box. The "saved" vendor cost is repaid with interest in the form of extended MTTR (Mean Time to Recovery) and organizational complexity.

## The Prediction

As the Authorization Gap widens, a reversal in how engineering talent is valued becomes probable.

In the near term, the "10x generator" will be celebrated. But as the latent costs of unverified code begin to surface in the form of reliability regressions and security incidents, the premium will shift. **The most valuable asset in an AI-accelerated organization will not be the ability to generate output, but the ability to authorize it.**

Staff and Principal roles will increasingly decouple from production entirely, moving toward a pure "Judgment Capital" model. Their primary function will be to act as the human circuit-breakers for a system that is producing more than it can perceive.

## Where the Model Fails

The Authorization Gap is less acute in low-stakes environments. If you are building a marketing landing page or a prototype, the cost of an authorization failure is low. In these contexts, velocity is the correct metric to optimize for.

The model also assumes that automated verification (testing, static analysis) cannot scale to meet the generation speed. If AI tools become significantly better at *verifying* code than at *generating* it, the gap might close. However, current trends suggest that generation is easier to automate than the nuanced judgment of "fit for purpose" and "long-term maintainability."

## Closing the Loop

The fundamental measurement problem remains. Organizations cannot optimize for what they cannot perceive. Velocity is legible; authorization depth is not.

Until the "audit debt" created by generated code becomes visible on a dashboard, the incentive structure will continue to favor the fire hose. The gap will continue to grow until the first major structural failure forces a recalibration of what "productivity" actually means in an age of infinite output.
80 changes: 80 additions & 0 deletions src/content/beyondthecode/the-maintenance-of-bespoke-tools.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,80 @@
---
title: "The Maintenance of Bespoke Tools"
date: 2026-02-24
description: "Why AI-enabled 'vibe-coding' is reviving Shadow IT and creating a long-term maintenance cliff for engineering organizations."
author: "Ganesh Pagade"
draft: false
---

<p class="drop-cap">The manager was tired of waiting for the platform team to prioritize a new deployment dashboard. He asked a Senior Engineer to spend an afternoon with an AI agent. By the end of the day, they had a custom React app that visualized the entire CI/CD pipeline. It worked perfectly. The team stopped complaining. The platform team's backlog remained untouched.</p>

On paper, this is a productivity win. The team unblocked themselves without consuming centralized resources.

**But in a political organization, every bespoke solution is a future maintenance liability disguised as a present-day efficiency.**

## The Revival of Shadow IT

For decades, IT departments fought "Shadow IT"—the practice of departments buying their own SaaS tools or building their own Excel-macros-turned-databases outside the purview of central governance. Governance won because building software was hard. You needed a budget, a server, and a developer.

AI-enabled "vibe-coding" has lowered the barrier to entry so far that Shadow IT is no longer a budgetary line item; it is a Tuesday afternoon activity.

**We are entering the era of Disposable Software.** If an engineer can build a custom CRM bridge or an incident-response bot in three hours, the hurdle for "build vs. buy" collapses. Why wait six months for a vendor review when you can prompt a solution into existence before the next standup?

## The Spatial Locality of Value

The value of these bespoke tools is spatially local. They solve a specific problem for a specific group of people at a specific point in time.

The Platform Team optimizes for "Organizational Legibility." They want every team to use the same deployment pipeline, the same monitoring stack, and the same architectural patterns. This makes the organization manageable at scale.

The individual Engineer optimizes for "Task Velocity." They want the dashboard *now* so they can finish their feature.

When an engineer vibe-codes a bespoke solution, they are trading Organizational Legibility for Task Velocity. The individual wins, but the organization loses a piece of its shared infrastructure. Multiply this by a hundred teams, and the organization becomes a thicket of incompatible, unmonitored, and undocumented "helper apps."

## The Maintenance Cliff

The "disposable" nature of AI-generated software is an illusion. Software is only disposable if it can be turned off without consequence.

In practice, these bespoke tools often become critical path. The "quick script" that syncs Jira to a custom Slack bot becomes the primary way the Director tracks project status. The "temporary" dashboard becomes the only way the team knows if a deployment failed.

**The debt is deferred, not avoided.**

The maintenance cliff appears when the original author leaves the company or rotates to a new project. The Staff Engineer who inherits the team finds themselves responsible for a suite of apps they didn't build, written in a "vibe-coded" style that prioritizes immediate function over long-term maintainability.

They cannot easily modify these tools because the "intent" of the code was often undocumented—it lived in the context window of the AI agent for three hours and then vanished. To change a feature, they must effectively re-prompt the entire app, hoping the AI doesn't introduce a regression that the author didn't have tests for.

## The VP's Dilemma

From the perspective of a VP of Engineering or a CFO, this looks like a cost-saving miracle. "We didn't need to spend $50k on a vendor license because our engineers built it themselves."

But the balance sheet only captures the *purchase* cost. It does not capture the "Latent Maintenance Cost"—the hours spent by high-salaried Staff Engineers three years from now trying to secure a legacy AI-generated app that has become a critical vulnerability.

The VP sees the immediate headcount efficiency. They do not see the "Complexity Tax" being levied on the organization's future.

## Surface Rationality

This is not a story of "lazy" engineers or "incompetent" managers. Both actors are behaving rationally within their incentive structures.

The Engineer is rewarded for unblocking their team. "Vibe-coding" a solution is the fastest way to earn that reward.
The Manager is rewarded for hitting deadlines. Skipping the vendor review process is a rational way to accelerate.

The conflict is structural. The organization needs standard, legible systems to survive at scale. The individual needs fast, bespoke systems to survive the current sprint. AI has simply given the individual a more powerful weapon in this tug-of-war.

## The Prediction

As vibe-coding becomes the default mode for internal tooling, the emergence of a new organizational layer becomes likely: the **Complexity Auditor**.

Their job will not be to write code, but to hunt down and "de-commission" bespoke AI-generated tools that have outlived their usefulness but refused to die. They will be the organizational antibodies, fighting the spread of Shadow AI to preserve the legibility of the system.

Organizations that fail to develop these antibodies will eventually find their "Core Business" work slowed to a crawl by the sheer friction of maintaining a thousand "simple" bespoke tools.

## Where the Model Fails

The "Bespoke Tool" model might succeed if AI agents become sophisticated enough to maintain *any* codebase, regardless of how it was originally written. If an agent can "understand" a legacy React app as easily as a human can, the maintenance cliff might not materialize.

However, this assumes that AI capability will stay ahead of the complexity generated by AI. It is a race between the speed of entropy and the speed of the agent. History suggests that software entropy is a remarkably resilient force.

## Closing the Idea

The tension between local productivity and global legibility is as old as engineering itself. AI hasn't changed the tension; it has merely increased the frequency of the trade-off.

We are currently in the "honeymoon phase" where the productivity gains are visible and the maintenance costs are still latent. The success of an engineering organization in the next five years will be determined not by how many tools they can prompt into existence, but by how many they can afford to keep.
X Tutup