Narrative · ▲ at peak

More code, but not more delivered software

All narratives · 22 supporting stories · 0 from Hungary

„AI agents multiply written code, but the volume of finished, released software and delivery speed do not increase as a result.”

Counter-narrative

Enthusiastic users and tool advocates say the gain is real, only the measurement lags, and delivery will speed up once processes are reorganized. This is mainly supported by developers' own time-saving estimates and reports of faster coding. According to one analysis, METR's follow-up even reversed the sign of the earlier slowdown.

Carriers

Harvard researchers (Jellyfish data), METR, NBER and Wharton researchers, GitClear, Atlassian and Jellyfish reports, LeadDev, The New Stack

Analysis

The narrative holds that AI accelerates coding in the development chain, while the gains are lost during review, testing, debugging, and operations. Data from Harvard, NBER, Wharton, GitClear, and Jellyfish point in the same direction. According to NBER, coding activity increased by 240% with agents, while GitClear measured an 81% increase in duplication alongside a 25% rise in output. A Hungarian report states that nearly 75% of Google's code is now written by AI, yet meaningful acceleration in development has not materialized. It is in the interest of vendors of review, debugging, and validation tools (Qodo, Undo, NeuBird) for the bottleneck to shift to the end of the chain, as that is where they sell. The weak point is that delivery is often measured by proxies (PR count, commits, throughput). Furthermore, Jellyfish data covering 1,300 companies shows that agents affect 27.7% of PRs at the median company, while at nine-tenths of companies, the share of PRs opened by agents is 0.0%, meaning agent-based work organization remains rare. It is debated whether this is a transitional state or a permanent limit.

Evidence: moderate — There are several large-sample measurements (718 companies, more than 500 thousand GitHub developers, 623 million changes), and they show growth in code volume alongside unchanged or slower delivery. However, delivery is mostly measured by proxy (PRs, commits, lead time). The Harvard study and the Jellyfish report rest on the same data source, and several sources are vendor surveys (Atlassian, Qodo, Undo). On top of that, METR's follow-up gave a signal in the opposite direction.

What's at stake: Whether corporate spending on AI coding translates into measurable business results or leads to budget cuts.

Narrative sub-threads

The narrative splits into 4 branches; a story may belong to several branches, and some may not appear in the main narrative's list.

▲ at peak · evidence: strong

Large-sample measurements: more code, no more delivery

Studies based on corporate and repo data show that AI tools and agents increase code output and coding activity by 25–240 percent. The volume of released software and development speed do not increase from this: in the Harvard study, pull request lead time increased by 14 percent.

This branch does not rely on self-reporting or case studies. It measures and quantifies the gap between output and delivery using data from hundreds of companies, hundreds of thousands of GitHub developers, and hundreds of millions of code changes.

7 stories · 0 of them Hungarian → lasting · evidence: moderate

METR debate and unreliable self-assessment

METR's experiments do not provide a clear picture of acceleration. The first study measured a 19% slowdown, while the follow-up result reversed the sign. Between 30–50% of developers invited to the follow-up study declined to perform AI-free comparative tasks. Self-assessment is an unreliable metric for acceleration.

In contrast to the telemetry branch, the subject here is the experimental methodology: the sign reversal between slowdown and speedup, withheld tasks, and the gap between perceived and measured speed.

4 stories · 0 of them Hungarian → lasting · evidence: moderate

Code review as a new bottleneck

Time gained in implementation is lost in code review and validation because AI generates more code than developers can follow.

It examines the single mechanism of delivery delays: review. Evidence comes from industry reports, a study of 200 people, practical cases, and expert analysis. The burden is carried by reviewers, mainly seniors.

9 stories · 0 of them Hungarian ↗ rising · evidence: moderate

Debugging, rework and deployment instability

AI code shifts the burden to debugging, testing, and human post-work. Undo research shows debugging consumes 42% of the workweek. A Harness report cites a 69% deployment error rate. One report states agents require human post-work for three-quarters of tasks. A survey indicates release cycles are slowing.

It is not about review, but the phase after review: bug fixing, testing, and live stability. The main evidence is a tool-vendor survey; the actors are testing and QA teams, as well as operations.

6 stories · 0 of them Hungarian

Sub-thread stories

Large-sample measurements: more code, no more delivery

Carriers: Harvard (Chen–Stratton, Jellyfish data), NBER, McKinsey and Wharton, GitClear, MIT

Sep 2, 2026MIT study: AI coding tools produce more code, but not more releasesCTOMeasure speed gains with release metrics, not lines of code.tech leadThe bottlenecks are review, testing and integration; that is where to strengthen the process.QA testerMore generated code means more verification work, so prioritize test automation.Sep 14, 2026GitClear: AI coding brings 25% more output and 81% more duplicationsenior developerDuplicated code should be handled through refactoring and review before it becomes technical debt.tech leadDuplication and maintenance burden should be measured alongside the impact metrics for AI use.CTOMaintenance cost should be included in the business case next to the increase in output.Sep 15, 2026McKinsey and Wharton: more AI-written code does not automatically bring business resultssenior developerReview and deployment are the bottleneck, so AI support should be applied there too.tech leadTeams should measure shipped features and lead time rather than activity metrics.CTOThe return on AI investment should be shown through business results, not the volume of code.Sep 28, 2026NBER study: coding activity up 240% with agents, but delivery bottlenecks remaintech leadMeasure the whole delivery pipeline (review, testing, release), not just the volume of code.CTOFaster coding does not automatically turn into faster product delivery.Oct 9, 2026Studies: AI agents write more code but do not speed up deliveryCTOCode volume is not a productivity metric, so the return on AI adoption should be measured by delivery and defect metrics.tech leadTeam metrics should include closed issues and delivery time, not lines of code.senior developerThe extra code also has to be reviewed, so the review burden may grow.Oct 12, 2026Harvard study: AI agents write 30% more code but do not ship more softwaretech leadMeasure PR lead time and review capacity rather than lines of code, and set agent workloads accordingly.CTODo not base a productivity business case on lines of code generated, but on features delivered and lead time.senior developerExpect more review work to land on you because of agents; agree PR size limits in advance.Oct 19, 2026Harvard study of 718 firms: AI produces more code but does not speed up deliveryCTOMeasure the return on AI investment across the whole delivery process (review, testing, release), not by the volume of generated code.tech leadReview and testing capacity must be expanded, because faster code generation causes congestion there.QA testerWith the increased volume of code, test automation and risk-based testing gain priority.

METR debate and unreliable self-assessment

Carriers: METR, Anthropic, Carnegie Mellon researchers

Code review as a new bottleneck

Carriers: Atlassian, Qodo, Gergely Orosz, Sourcegraph

Aug 4, 2026Agentic AI rewrites code review economics as Rootly drops its small-PR ruletech leadRules on PR size and review process need rethinking for the volume of change agents produce.senior developerSplit large AI PRs into smaller, stacked changes before review.QA testerReviews can rely on automated evidence (tests, checks) instead of line-by-line reading.Aug 17, 2026Articles urge risk-tiered rules as AI-generated code review burden growstech leadIntroduce a risk-tiered review rule so that large AI-generated diffs do not slip through without being reviewed.senior developerReduce the review load with smaller units of change and automated checks.QA testerTests and acceptance criteria will play a more prominent role in verifying AI-generated code.Sep 8, 2026AI speeds up software development, but the bottleneck shifts to review and operationsCTOProductivity should be measured against finished, live software, not the amount of code generated.QA testerThe larger volume of changes puts more testing load on QA.Sep 8, 2026AI broke code review: experts disagree on what should replace ittech leadThe review process needs redesigning: what humans check, what machines check, and the upper limit on PR size.senior developerReview becomes a core competence, as checking for errors is harder than writing the code.QA testerAutomated test coverage becomes the main line of defense against AI-generated changes.Sep 21, 2026After AI code generation, verification is the bottleneck: more commits, not proportionally more releasestech leadInvest in review capacity and automated verification, because writing speed is no longer the constraint.QA testerValidation gains in value, and testing and checking AI code become a central task.CTOMeasure impact by release frequency and defects, not by commit count.Sep 29, 2026200-person study: AI coding speeds up implementation but leaves harder engineering decisionssenior developerPlan more time for review and architectural decisions.tech leadDo not raise expectations automatically in proportion to faster implementation.Oct 9, 2026AI-generated code is burning out senior engineers in code review, analysis sayssenior developerRecurring review feedback should be recorded as rules so that automation can check it.tech leadPlan how review load is shared and how AI reviewers are introduced before reviewers burn out.Oct 28, 2026Atlassian: AI time savings are disappearing into code reviewtech leadReview is the bottleneck, so add capacity or automation there.senior developerMore AI-generated code means more review, so reviewer time must be budgeted.Nov 2, 2026AI code review is the next bottleneck: more code, fewer lines readtech leadThe review process needs redesigning, because the volume of code from agents exceeds human review capacity.senior developerYou have to decide what to review line by line and what to leave to automated checks.

Debugging, rework and deployment instability

Carriers: Undo, Harness, Tech Field Day

Sep 15, 2026Harness report: AI speeds up development, but deployment failure rate is 69%QA testerWith more code being produced, test coverage and regression filtering are critical.CTOAI-driven acceleration should be tracked together with release-quality metrics.Sep 28, 2026AI coding imposes a 'quality tax' on bank testingQA testerPrepare for a heavier testing load and for targeted regression checks of AI-generated changes.Sep 29, 2026Undo research: AI agents push debugging to 42% of the workweektech leadSchedule dedicated debugging time when AI speeds up delivery.senior developerInvest in debugging and observability tools, because the need to verify AI-written code is growing.Oct 7, 2026Survey: AI-generated code shifts the bottleneck to debugging, slowing releasesQA testerTesting and debugging may face a heavier load, so capacity should be planned in advance.tech leadWhen measuring release cycles, time spent on debugging should be tracked separately.Oct 12, 2026Quality debt: AI-accelerated development brings production outages and testing gapsQA testerWith the flood of AI-driven changes, risk-based, automated and exploratory testing gain value.tech leadIntroduce quality gates that AI-generated code cannot bypass.Nov 8, 2026AI coding agents still need human rework on three-quarters of taskstech leadDo not treat agents' output as finished; build rework time into your estimates.senior developerTaking over and fixing agents' work is the main part of the daily job, so review skills are valuable.

Google search interest

Hungary

Topic and search term	Weekly	Average
More code, but not more delivered softwareAI coding productivity	no measurable search	0
Large-sample measurements: more code, no more deliverydeveloper productivity measurement	no measurable search	0
METR debate and unreliable self-assessmentMETR AI study	no measurable search	0
Code review as a new bottleneckcode review	no measurable search	0
Debugging, rework and deployment instabilityAI code errors	no measurable search	0

World

Topic and search term	Weekly	Average
More code, but not more delivered softwareai coding productivity		18.4
Large-sample measurements: more code, no more deliverydeveloper productivity metrics		0.6
METR debate and unreliable self-assessmentmetr ai study		1.8
Code review as a new bottleneckcode review bottleneck		0.2
Debugging, rework and deployment instabilityai generated code bugs	no measurable search	0

Google Trends relative search index over the period studied, weekly average. A narrative and its sub-threads appear in a single comparison: 100 is the group's most-searched day, and the figures can only be compared within the group. The search term was chosen by machine analysis; it measures how much the term is searched, not how widespread the narrative is.

Who is affected and how

CTO: The return on AI spending should not be measured by generated lines or commits, but by lead time, change failure rate, and the number of released features. Before expanding licenses, it is advisable to run an internal measurement with a control group.
tech lead: The team's bottleneck is the PR queue, CI, and testing; capacity must be reallocated there, and PR turnaround time tracked weekly.
senior developer: Self-estimated time savings are unreliable. It is worth consciously measuring on a few tasks where time is actually spent (debugging, review, integration).
QA tester: The growing volume of code is overwhelming testing. This strengthens the position of test automation and validation competencies, which should be built up with a risk-based testing strategy.

Hungary

Domestic footprint during the period under review is minimal. There is no [HU]-marked story in the list; the topic appears only in Hungarian-language reports on global results (Google 75%, agents' post-work requirements, Economist coverage). There is no data on domestic delivery metrics.

Weekly trend

Supporting stories per week

Turning points

Aug 17, 2026METR follow-up study reverses the 19% slowdown resultAn August 17 analysis found that METR's follow-up reversed the sign of the first experiment's 19% slowdown (16 developers, 246 tasks), so the old figure is outdated.Oct 12, 2026Harvard study: AI agents write 30% more code but do not ship more softwareFiona Chen and James Stratton's Harvard study covering 718 companies (approximately 30% more code, an unchanged amount of finished software, 14% longer PR turnaround) supported the paradox with large-sample research results.Oct 26, 2026METR follow-up: many developers withheld tasks; self-reported speedups are unreliableIn the METR follow-up study, 30–50% of invited developers did not submit tasks for the AI-free comparison. Reports indicate that self-estimation is therefore a poor metric for acceleration.

Related narratives in the News magazine

The AI investment bubble AI replaces white-collar work

Supporting stories

Aug 4, 2026Research on AI coding's impact: slower debugging, an inconclusive METR study, mythstech leadMeasure speed gains across the whole delivery chain (review, debugging), not by code-generation speed.CTOBefore planning around productivity promises, you need data measured on your own team.senior developerTo keep debugging skills sharp, practise without an agent on critical parts.junior developerRelying on AI can reduce learning, so practise understanding code deliberately.Aug 17, 2026METR follow-up study reverses the 19% slowdown resultCTOInterpret productivity claims by task type and codebase maturity rather than relying on a single number.tech leadMeasure the effect on your own team before committing to a roadmap built on AI tools.Sep 2, 2026MIT study: AI coding tools produce more code, but not more releasesCTOMeasure speed gains with release metrics, not lines of code.tech leadThe bottlenecks are review, testing and integration; that is where to strengthen the process.QA testerMore generated code means more verification work, so prioritize test automation.Sep 8, 2026AI speeds up software development, but the bottleneck shifts to review and operationsCTOProductivity should be measured against finished, live software, not the amount of code generated.QA testerThe larger volume of changes puts more testing load on QA.Sep 14, 2026GitClear: AI coding brings 25% more output and 81% more duplicationsenior developerDuplicated code should be handled through refactoring and review before it becomes technical debt.tech leadDuplication and maintenance burden should be measured alongside the impact metrics for AI use.CTOMaintenance cost should be included in the business case next to the increase in output.Sep 15, 2026McKinsey and Wharton: more AI-written code does not automatically bring business resultssenior developerReview and deployment are the bottleneck, so AI support should be applied there too.tech leadTeams should measure shipped features and lead time rather than activity metrics.CTOThe return on AI investment should be shown through business results, not the volume of code.Sep 21, 2026After AI code generation, verification is the bottleneck: more commits, not proportionally more releasestech leadInvest in review capacity and automated verification, because writing speed is no longer the constraint.QA testerValidation gains in value, and testing and checking AI code become a central task.CTOMeasure impact by release frequency and defects, not by commit count.Sep 28, 2026NBER study: coding activity up 240% with agents, but delivery bottlenecks remaintech leadMeasure the whole delivery pipeline (review, testing, release), not just the volume of code.CTOFaster coding does not automatically turn into faster product delivery.Sep 29, 2026Undo research: AI agents push debugging to 42% of the workweektech leadSchedule dedicated debugging time when AI speeds up delivery.senior developerInvest in debugging and observability tools, because the need to verify AI-written code is growing.Sep 29, 2026200-person study: AI coding speeds up implementation but leaves harder engineering decisionssenior developerPlan more time for review and architectural decisions.tech leadDo not raise expectations automatically in proportion to faster implementation.Oct 7, 2026Survey: AI-generated code shifts the bottleneck to debugging, slowing releasesQA testerTesting and debugging may face a heavier load, so capacity should be planned in advance.tech leadWhen measuring release cycles, time spent on debugging should be tracked separately.Oct 9, 2026Studies: AI agents write more code but do not speed up deliveryCTOCode volume is not a productivity metric, so the return on AI adoption should be measured by delivery and defect metrics.tech leadTeam metrics should include closed issues and delivery time, not lines of code.senior developerThe extra code also has to be reviewed, so the review burden may grow.Oct 12, 2026Harvard study: AI agents write 30% more code but do not ship more softwaretech leadMeasure PR lead time and review capacity rather than lines of code, and set agent workloads accordingly.CTODo not base a productivity business case on lines of code generated, but on features delivered and lead time.senior developerExpect more review work to land on you because of agents; agree PR size limits in advance.Oct 12, 2026Productivity perception gap: METR studies and developers' dependence on AItech leadDo not rely on developers' impressions; measure in a controlled way before deciding on tool adoption.senior developerNotice which tasks you only want to do with AI now, and keep your core skills in practice.Oct 16, 2026Nearly 75% of Google's code is now written by AI, but speed-up lagsCTOThe share of AI-written code is not the same as delivery speed; measure lead time.tech leadIncrease review and testing capacity as well if more code is being produced.Oct 19, 2026Harvard study of 718 firms: AI produces more code but does not speed up deliveryCTOMeasure the return on AI investment across the whole delivery process (review, testing, release), not by the volume of generated code.tech leadReview and testing capacity must be expanded, because faster code generation causes congestion there.QA testerWith the increased volume of code, test automation and risk-based testing gain priority.Oct 20, 2026Atlassian survey: over 95% of developers use AI, but delivery friction remainstech leadTeams should identify the steps in their processes where work piles up after coding gets faster.CTOAlongside investment in AI tools, plan for improving the delivery process as well.Oct 26, 2026METR follow-up: many developers withheld tasks; self-reported speedups are unreliabletech leadDo not base productivity on self-reports, because METR says they are biased.CTOBack up the return on AI investment with objective measurements.Oct 26, 2026Jellyfish: agents touch 27.7% of pull requests, while "software factories" remain raretech leadAgents currently assist human work more than they open PRs on their own, so tailor the review process accordingly.CTOFull automation is still rare, so the PR share is a realistic target to measure.Oct 28, 2026Atlassian: AI time savings are disappearing into code reviewtech leadReview is the bottleneck, so add capacity or automation there.senior developerMore AI-generated code means more review, so reviewer time must be budgeted.Nov 2, 2026AI writes 75% of Google's new code as developer distrust risestech leadAs the share of AI-generated code grows, code review culture and test coverage need to be strengthened in advance.senior developerReview has become your main job, so it is worth developing methods for checking AI-generated code.Nov 8, 2026AI coding agents still need human rework on three-quarters of taskstech leadDo not treat agents' output as finished; build rework time into your estimates.senior developerTaking over and fixing agents' work is the main part of the daily job, so review skills are valuable.

Machine analysis based on story titles and summaries, with a verification pass. The source of the facts is the article linked at each story.