@svpino: We haven’t figured out how to scale from producing stuff to shipping stuff: • 2,450% increase in lines of code • 240% i…
Summary
Sonar CEO Tariq Shakir’s talk points out that AI agents have caused code output to surge (+2,450% in lines of code, +240% in commits), yet releases have grown only 30% — the “scaling from production to delivery” problem remains unsolved. He proposes a “code confidence formula” built by multiplying intent, guardrails, verification, and evidence, helping software factories deliver confidence rather than blind trust.
View Cached Full Text
Cached at: 11/03/26, 02:19 AM
We haven’t figured out how to scale from producing stuff to shipping stuff:
• 2,450% increase in lines of code • 240% increase in commits • but only 30% increase in releases
This video is really cool:
https://t.co/Oi1reBDNL8 https://t.co/8Xs0GyFJ7n
TL;DR: Sonar CEO Tariq Shakir argues that AI agents have caused code output to skyrocket (+2,450% lines of code, +240% commits) while delivering only +30% in releases — the “from production to shipping” problem remains unsolved. He proposes a “Code Confidence Formula” built from clear intent, guide rails, independent validation, and evidence — all multiplied together — to help software factories deliver confidence rather than blind trust.
From the Toyota Factory to the Software Factory
The talk opened with a real-world factory story. In the 1980s, car parts had visible gaps, and even a good car couldn’t perform exactly as intended. Then Toyota (especially Lexus) ran an ad: a ball bearing rolling smoothly across a body panel seam. That was revolutionary at the time — it proved cars could be perfectly assembled, shifting the mindset from “ship volume first, rework later” to “get it right the first time, zero rework, customer satisfaction.”
Tariq Shakir summarized the elements of that factory philosophy that still hold true today:
- Prioritize quality over volume.
- Fix the system, not just individual symptoms.
- Design for manufacturability — instead of designing wheels, reservoirs, and every part first, then hoping they fit together.
- Solve problems the moment they appear — don’t kick the can down the road.
- Empower everyone — anyone can stop the production line, not just the boss.
Software Development Has Inherited These Principles, but Agents Changed the Game
Modern software development absorbed these ideas: Agile, Kanban, error budgets (in the SRE world), continuous integration — all born from the concept of “how to build an effective factory.” So far, they’ve worked very well.
But agents are changing the rules — a study published last month on implementing asynchronous agents across the software development lifecycle found:
- 2,450% growth in lines of code
- 240% growth in commits
- Only 30% growth in releases
Code creation has become easy and abundant. Sonar just completed a “State of Code” developer survey coming out next month: right now 5% of surveyed developers are running agentic software factories, 34% expect to do so within a year, and 39% in financial services.
At the same time, the pitfalls are emerging: agents can help you build things — and they’re also very good at helping you hack, attack, find vulnerabilities, and surface quality issues.
Quoting Satya Nadella: Engineered Containment and Governance Are Needed
While preparing his demo over the weekend, Shakir read a post by Satya Nadella on X about the internal risks of in-house models, which included this point: there needs to be an engineered approach to containment and governance — built around powerful deterministic systems, human controls, and reliable operational procedures.
This echoes his core judgment: software development is undergoing a paradigm shift, and we need to build systems that embed confidence and quality into agentic software factories.
Five Limitations
1. Human-in-the-loop doesn’t scale
Every company talking about their AI governance strategy says, “Don’t worry, we have a human in the loop.” But with code lines up 2,450%, commits up 240%, and releases up only 30% — we haven’t figured out how to scale from producing to shipping.
What’s even more critical is cognitive surrender: a study from a few months ago examined how often humans agree with AI. The results: when AI is right, people agree 93% of the time; when AI is wrong, they agree four out of five times as well. So the question isn’t “put a human in the loop to go faster” — it’s that this approach simply doesn’t work in the agentic era.
2. Agents perform poorly in large, complex codebases
A code translation study found:
- At 10,000 lines of code, agents achieve 91% accuracy
- At 50,000 lines, accuracy collapses
Another study measured the tokens required to complete the same task in a large, complex codebase — the difference was 30x (not 30%), because agents spend enormous amounts of tokens understanding the codebase structure and figuring out how to navigate it. Shakir used an analogy: traveling from Los Angeles to New York — if you randomly start in White Plains, you’ll get there quickly; if you start in Albuquerque, it takes a long time and a lot of cost. Agents are the same.
3. Errors accumulate
Research from Meter (available on their site) measures an agent’s ability to complete human-level tasks, using “can it complete a task that takes X hours” as the metric. Claude Mythos preview is good at completing tasks that take humans 16 hours — but that’s at 50% accuracy. When the accuracy requirement is raised to 80%, the curve flattens dramatically, because errors accumulate over long-running tasks: if each step has 95% accuracy, by step 20 accuracy has dropped to just 30–36%.
4. LLMs are unreliable, reward-seeking judges
“Just ask the LLM?”
- A paper published this month by the PayPal AI Research Lab found that across different models and scenarios, 7% to 17% of verdicts changed across what were supposed to be “identical replications” — ask the same question, and there’s a 7–17% chance of getting a different answer.
- Meta’s research shows that simply asking “Are you sure?” causes the LLM to change its mind 7% to 34% of the time (depending on the model).
Ask the same AI judge twice, and you may get two different answers.
5. Time to exploitation is near zero — response capability faces enormous challenges
This point is self-evident to everyone in the room and won’t be elaborated on. Prevention must be engineered into the design of the software factory.
The Code Confidence Formula: Four Elements, Multiplied
Sonar wrote down what they call the “Code Confidence Formula.” The goal isn’t blind trust — it’s enabling you to tell executives, customers, shareholders, the CEO, and the board: “Yes, I’m confident this factory is operating well.”
- Intent — if you don’t know what to build, you can’t tell the agent what to build.
- Guide rails — deliberately said “guide rails” rather than “guard rails”: guard rails keep you from going off the road; guide rails help you stay on it. Agents need small nudges to steer and manage them.
- Validation — must be independent and thorough; problems must be fixed, not deferred.
- Evidence — trust but verify; trust but understand it has proof: consistent, auditable, reproducible.
The four are multiplied: if any one is zero, confidence goes to zero — doing all of this without evidence still results in zero.
Real Customer Outcomes
Intent and Guide Rails: Sonar Vortex
Sonar Vortex provides intent guardrails and begins validation within the agentic loop, then nudges the agent forward. When given this information to run long tasks in large, complex codebases over extended periods:
- 36% reduction in token costs — place it in White Plains, near the answer, help it understand the map, so it doesn’t waste tokens on things it already knows.
- 92% fewer issues in commits.
Validation: SonarQube
Sonar started as a validation company. Independent validation must cover the full scope of issues, both precise and efficient: use CPU when CPU works, use GPU when GPU is needed — a deterministic hybrid approach plus an agentic hybrid approach forms what they see as the “gold standard of validation.”
Sonar recently surveyed over 1,000 customers, asking how frequently they encounter problems with agentic coding. The conclusion: SonarQube users are significantly more likely to see a very positive impact than non-users —
- 81% more likely to see a very positive impact on code security
- 34% more likely to see a positive impact on time to market
- 76% more likely to see a positive impact on failure frequency
Evidence: Ship with Evidence
You must pass quality gates — whether internal controls and IT general controls (ITGC), MISRA for automotive, or any other standard — generating the evidence you need to sleep at night and the confidence to move forward.
Conclusion: Code Is Abundant, Confidence Is Scarce
Code is abundant, confidence is scarce. Our goal — and what we believe everyone should be aiming for when thinking about the future of software development — is: how do we deliver confidence? The factory of the future must be able to deliver confidence.
Source: https://www.youtube.com/watch?v=LQW_Wvqhp04
Similar Articles
@rohanpaul_ai: MIT study. Code volume surges by 300%, but output increases by only 30%: The AI dividend meets an awkward reality. They…
An MIT study of over 100,000 GitHub developers finds that AI coding tools increase code volume by up to 300% but only boost shipped software by 30%, highlighting bottlenecks in human review and integration.
@omarsar0: Nice report on agents beyond code generation. Here is why it matters: Coding agents raise how much code gets written. T…
A report synthesizes field studies to argue that while AI coding agents increase code output, gains shrink in shipping reliable software due to bottlenecks in review, integration, and cost management.
@thinkszyg: The AI Coding Speed Paradox: Coding 48% Faster, Review 6x Slower. How to Rebuild the Review Process? SD Times Analyzes Data from 250,000 Developers: AI Boosts Coding Speed by 48-58%, But AI-Generated PRs Get Stuck in Review for 4-6x Longer…
The article points out that AI coding increases coding speed by 48-58%, but code review time increases by 4-6x, and security vulnerabilities also rise. It proposes a three-step plan to rebuild the review process: AI pre-review, focusing on architectural decisions, and using Microsoft's open-source ASSERT framework for behavioral verification.
@saranormous: https://x.com/saranormous/status/2064510215056400652
Despite rapid advances in AI coding agents like Devin, which have dramatically increased code writing and shipping, the article argues that the most valuable aspects of software engineering remain illegible to benchmarks and require human judgement and organizational coordination that cannot be easily automated.
@svpino: For many teams, coding is no longer the bottleneck, but you still have to decide what to build using agents. When execu…
A discussion from the State of the AI SDLC Digital Summit highlights how AI is accelerating software development, shifting the bottleneck from coding to product decisions.