AI has made code creation abundant, but releases have grown much more slowly, so software factories need new ways to control quality.
2
Human review does not scale because people often accept incorrect AI output, while agents lose accuracy on large codebases and long tasks.
3
Sonar's code confidence formula combines clear intent, guide rails, independent verification, and repeatable evidence.
Summary
Tariq Shaukat compares agentic software development with the factory changes associated with Toyota and Lexus. Earlier software practices borrowed ideas such as quality at the source, continuous integration, and empowering people to stop the line. AI agents disrupt that model by making code cheap and plentiful while leaving release capacity, review, and confidence behind. Shaukat describes five problems: human review does not scale, agents struggle with large codebases, errors compound over long tasks, LLM judges can change their verdicts, and the time to exploit a vulnerability is nearly zero. His answer is a software factory built around four multiplied components: clear intent, guide rails, independent verification, and evidence that is consistent, auditable, and repeatable. He presents Sonar customer results for token costs, commit issues, security, time to market, and outages, while arguing that confidence must be designed into the factory rather than added at the end.
Toyota-style factory thinking puts quality into the system
Shaukat begins with Toyota and Lexus cars from the 1980s, using a ball bearing rolling across the seams of a car to illustrate precise assembly. The factory lesson is to prioritize quality over volume, fix the system rather than isolated symptoms, design parts so they fit together, solve problems when they appear, and let anyone stop the production line. He connects these ideas to software practices such as Agile, Kanban, error budgets, and continuous integration. The shared principle is to build a production system that delivers quality instead of relying on final repair.
Agents make code abundant while release capacity lags behind
Research on asynchronous agents showed a 2,450% increase in lines of code and a 240% increase in commits, while releases increased by only 30%. Shaukat says this creates a gap between producing code and shipping software. He also previews a developer survey in which 5% of respondents were already running agentic software factories, while 34% expected to do so within a year. The new factory therefore needs controls for both production and the security risks that come with agents that can also find vulnerabilities and attack systems.
Human review fails when people surrender judgment to AI
Shaukat argues that keeping a human in the loop cannot solve the scale problem. The same study that found code creation rising by 2,450% found releases rising by only 30%. He also cites research on cognitive surrender. People agreed with AI 93% of the time when it was right, but they also agreed four out of five times when it was wrong. Faster review is not enough if reviewers accept incorrect output. Agentic factories need automated checks and other controls rather than relying on human approval for every change.
Large codebases make agents spend effort finding their way
Agents can be very capable but still struggle to understand a large, complex codebase. In one code translation study, they were correct 91% of the time on a 10,000-line codebase, while accuracy collapsed at 50,000 lines. Another study found a 30x variance in the number of tokens needed for the same task. Shaukat attributes this to the agent spending tokens understanding the codebase's structure and working out how to move through it. The starting point matters, much like choosing a nearby or distant starting location for a trip.
Long agent tasks turn small errors into poor outcomes
A study from METR measured whether agents could complete tasks that take humans different amounts of time. Claude Mythos Preview could complete tasks taking a human 16 hours, but that result used a 50% accuracy threshold. At an 80% threshold, the curve flattened substantially. Shaukat explains that errors compound across steps: with 95% accuracy at each step, a 20-step task has only 36% accuracy. Long-running software tasks therefore need checks during execution rather than a single review at the end.
LLM judges can change their verdicts on the same case
Shaukat questions the idea that an LLM can simply judge whether another agent's work is correct. Research from the PayPal AI research lab found that verdicts changed between 7% and 17% of the time across identical replications, depending on the model and situation. A separate Meta study found that asking an LLM, 'Are you sure?' changed its answer between 7% and 34% of the time. These results make an LLM-only judging process unreliable. Verification needs methods that do not depend on a single variable model response.
Confidence depends on intent, guide rails, verification, and evidence
Shaukat's code confidence formula has four components. Intent defines what the team is trying to build. Guide rails give agents small nudges that keep them on the intended path. Verification must be independent, thorough, and connected to fixing issues. Evidence must prove what happened in a consistent, auditable, repeatable way. He describes the formula as multiplicative because if any component is missing, confidence falls to zero. The model is designed to produce outcomes that people can explain to executives, customers, boards, and shareholders.
Sonar's customer results connect guidance and verification to outcomes
Shaukat says Sonar Vortex supplies intent, guide rails, and verification during the agent loop. In long-running tasks on large codebases, Sonar observed a 36% reduction in token cost and a 92% reduction in issues in the commit. He then describes SonarQube's verification approach, which combines deterministic methods with agentic methods and uses CPUs or GPUs according to the task. In a survey of more than 1,000 customers, Sonar users were 81% more likely to report a very positive security impact from AI coding, 34% more likely to report a time-to-market impact, and 76% more likely to report an impact on outage frequency.
"Instead of building for volume and building and then fixing the old mindset, you could actually have a mindset focused on ship it right, on no rework, on customer happiness."01:10
Who should watch
You are building internal agent workflows and need a way to control quality without manually reviewing every generated change.
Your agents work on large repositories or long-running tasks, and token use, error accumulation, or inconsistent reviews are becoming problems.
You need to explain to security, operations, or executive teams what evidence supports shipping AI-generated code.