Harvard researchers Fiona Chen and James Stratton found a hard truth for teams shipping with AI coding agents: you get more code, more commits, and more pull requests, but not more finished software. Human code review becomes the bottleneck. That finding, covered by Ars Technica on October 9, 2026, matters for every India and USA agency that sells velocity without staffing review capacity.

I am Nitin Kumar, Lead Project Coordinator at Debugged Pro. This guide sticks to the Chen and Stratton results via Jellyfish analytics, adds a clearly labeled second lens from Demirer and colleagues, and maps next steps to the NIST AI Risk Management Framework. No invented quotes. No padded stats.

AI coding agents review bottleneck graphic for India and USA agencies — Feature graphic: AI coding agents raise code volume while review time grows. Original Debugged Pro illustration.

What the Chen and Stratton Study Measured

According to Ars Technica coverage by Kyle Orland, Chen and Stratton used aggregated Jellyfish analytics across about 300 million work events, more than 700,000 employees, and over 700 software firms from 2021 through March 2026.

They mixed direct AI usage signals with GitHub activity to mark when firms added AI coding assistants (human-led autocomplete style help) and AI coding agents (prompt-driven code that can write and submit more on its own). Then they ran difference-in-differences style checks before and after those rollouts.

By March 2026, about 95 percent of firms in the study had implemented AI coding agents. Adoption is no longer a niche experiment. The question is whether more agent code equals more shipped value.

AI Coding Agents Raise Code Volume, Not Shipping Rate

On raw production metrics, the agent effect is large and clear:

About 30 percent more lines of code after AI coding agents arrive.
About 20 percent more commits.
About 23 percent more pull requests.

Those numbers look great in a sprint dashboard. They look weaker when you ask what actually closed. Issue and Epic resolution rates in Jira-like trackers did not change in a statistically significant way. The researchers also did not find a clear compositional shift in the size or complexity of those tracked issues.

In plain agency language: the backlog still moves at roughly the same feature pace. The team just writes and opens more change sets to get there.

AI coding agents output versus shipping metrics chart for agencies — Chen and Stratton pattern via Ars Technica: more lines, commits, and PRs; longer review; flat issue resolution. Original Debugged Pro illustration.

Why Pull Request Review Becomes the Bottleneck

The missing time shows up in review. Average review process time from pull request submit to merge rose about 49 percent after AI coding agents entered the workflow.

More detail from the same Ars Technica summary of the study:

Share of PRs with changes requested nearly doubled.
Comments per pull request rose about 35 percent.
Share of workers doing code reviews rose about 14 percent.

Teams responded by spreading review work across more people. That is not free. Senior engineers in India delivery pods and USA client pods already carry design, estimation, and incident load. A larger review queue steals the same calendar that clients watch for release dates.

Employment did not move in a way the authors could pin on AI. Looking across Jellyfish and LinkedIn signals, they could not attribute significant employment changes to AI. You do not automatically hire fewer people when agents write more. You often ask the same people to review harder.

AI Code Review Helps Only at the Margin So Far

Could AI also clear the review queue? In this dataset, only a little. By March 2026, about 80 percent of measured firms used some form of AI code review. Yet AI agents wrote only about 23.3 percent of review comments and accounted for about 10.8 percent of pull requests. Humans still did most review work.

That split is useful for client conversations. Do not sell "AI writes and AI merges" as a finished operating model. Sell "AI drafts, humans gate, AI assists review" and price the human gate honestly.

A Second Study Lens: Demirer et al. on Writing vs Shipping

Treat this next block as a separate study, not the same sample. Demirer, Musolff, and Yang (NBER Working Paper 35275), summarized in a CEPR/VoxEU column, looked at GitHub and Microsoft usage across generations of AI coding tools. Chen and Stratton used Jellyfish firm analytics. Demirer and colleagues used developer-level GitHub and Microsoft records. Same theme. Different design.

Their task-level gains grow across tool generations:

Autocomplete: about 40 percent more commits.
Cumulative with sync agents: about 140 percent more commits.
Cumulative with async agents: about 180 percent more commits.

Higher in the shipping chain, gains shrink. Combining generations, they report about 50 percent more projects and about 30 percent more releases. For sync agents alone, a more than sevenfold rise in lines of code becomes about 65 percent more pull requests and only about 20 percent more releases.

Their model framing matches what agency leads already feel: the binding constraint is shifting from writing code to review, integration, test, and release. That is the same story Chen and Stratton tell with firm-level review timers.

What This Means for India and USA Agencies

Most clients do not buy lines of code. They buy closed Epics, stable releases, and predictable calendars. If your proposal still prices only "developer hours with AI copilots," you will underfund the review bottleneck and overpromise ship dates.

Common failure modes we see across India and USA delivery:

Agent surge without reviewer surge. PR volume jumps; merge SLAs slip.
Metric gaming. Dashboards celebrate commits while Jira Epics stall.
Offshore/onshore handoffs. Async comments pile up across IST and ET without a named merge owner.
Client demos of agent speed. Stakeholders see a green CI run, then wait weeks in review.

If you already publish on AI agent production trust for India USA agencies, pair that trust story with this delivery story. Clients need both: safe runtime and honest throughput math.

Rewrite the pitch before the next SOW

Split coding capacity from review capacity in the staffing plan.
Put PR submit-to-merge on the weekly scorecard next to commit counts.
Cap agent PR size so reviews stay readable.
Name a merge owner per time zone so India and USA shifts do not park the same PR for days.

Practical Playbook When You Roll Out AI Coding Agents

Use this as a delivery checklist. Assign owners before the next agent pilot goes client-facing.

Baseline before agents. Capture lines, commits, PRs, review cycle time, change-request rate, and Epic resolution for four weeks.
Set a review budget. If PR volume rises about 20 percent, plan reviewer hours before you turn agents loose on every ticket.
Prefer smaller agent diffs. Large AI-generated patches nearly double change requests in the Chen and Stratton pattern. Keep PRs reviewable.
Require human merge for production paths. AI review comments can assist. They should not auto-merge client-critical branches.
Track comments per PR and rework rate. A 35 percent comment jump is a signal, not noise.
Do not cut headcount on agent hype alone. The study found no significant AI-driven employment drop. Cutting reviewers deepens the bottleneck.
Train agents on team standards. Style guides, test rules, and forbidden patterns reduce first-pass reject rates.
Report shipping metrics to clients. Show Epic and release movement, not raw code volume.

Agency checklist to fix AI coding agents code review bottleneck — Agency checklist for AI coding agents: review capacity, PR size caps, NIST mapping. Original Debugged Pro illustration.

If the client needs process redesign, not just more seats, point them to software development consulting with a clear review-capacity workstream. Tool licenses alone will not fix a 49 percent longer review cycle.

Map the Bottleneck to NIST AI RMF

India and USA buyers often ask for a named framework. Tie AI coding agents to the NIST AI Risk Management Framework so security and delivery teams share one checklist:

Govern: who may enable agents on a repo, who owns merge rights, who reports review SLAs to the client.
Map: which services allow agent-authored PRs, which paths need dual review, and where AI review tools run.
Measure: PR submit-to-merge time, change-request rate, comments per PR, Epic resolution, and agent share of review comments.
Manage: PR size caps, human merge gates, reviewer surge plans, and pause rules when rework spikes.

This is risk language clients already accept. It also stops the false story that more generated code automatically means lower delivery risk.

Client Conversation Script You Can Reuse

Keep the first update short. Boards want facts they can check.

What the research shows: Chen and Stratton (Harvard, Jellyfish data through March 2026) saw about 30 percent more lines, 20 percent more commits, and 23 percent more PRs after AI coding agents, with no significant Issue/Epic resolution lift and about 49 percent longer review cycles.
What we measured here: our baseline review time, change-request rate, and Epic throughput for the last four weeks.
What we changed: reviewer staffing, PR size limits for agent output, and merge ownership across India and USA shifts.
What we will report next: submit-to-merge trend and Epic closures, not commit vanity metrics.

Browse more agency delivery notes on the Debugged Pro blog when you need adjacent topics on trust, incidents, and agent operations.

Key Takeaways

AI coding agents raise code volume. Chen and Stratton report about 30 percent more lines, 20 percent more commits, and 23 percent more PRs.
Shipping did not rise with code volume. Issue and Epic resolution did not move in a statistically significant way.
Human review is the bottleneck. Average PR submit-to-merge time rose about 49 percent; change requests nearly doubled; comments per PR rose about 35 percent.
AI review is still secondary. By March 2026, AI agents made about 23.3 percent of review comments and 10.8 percent of PRs, even as 80 percent of firms used some AI review.
A separate Demirer et al. study shows the same attenuation. Large commit gains shrink to roughly 50 percent more projects and 30 percent more releases across tool generations.
Agencies must staff and measure review. Price reviewer capacity, cap agent PR size, and report Epic and release outcomes to India and USA clients.

FAQ: AI Coding Agents and the Review Bottleneck

Do AI coding agents help teams ship more software?

In the Chen and Stratton Jellyfish study summarized by Ars Technica, firms saw more lines, commits, and pull requests after AI coding agents, but Issue and Epic resolution rates did not rise in a statistically significant way. Extra coding time was absorbed by longer human review.

How much slower did code review become?

Average review process time from pull request submit to merge rose about 49 percent after AI coding agents were introduced. The share of PRs with changes requested nearly doubled, and comments per PR rose about 35 percent.

Are Chen/Stratton and Demirer et al. the same study?

No. Chen and Stratton (Harvard) used Jellyfish firm analytics across 700-plus software firms. Demirer, Musolff, and Yang used GitHub and Microsoft developer records (NBER WP 35275). Both find coding gains that shrink at shipping stages, but they are separate designs.

Can AI code review remove the bottleneck?

Not yet at scale in the Chen and Stratton data. By March 2026, about 80 percent of measured firms used some AI code review, yet AI agents produced only about 23.3 percent of review comments and 10.8 percent of PRs. Humans still did most review work.

What should India and USA agencies change first?

Baseline review cycle time, fund reviewer hours before agent rollouts, cap agent PR size, keep human merge gates on production paths, and report Epic and release outcomes instead of raw commit counts. Map controls to NIST AI RMF language for client risk reviews.

About the author: Nitin Kumar is Lead Project Coordinator at Debugged Pro. He works with India and USA agency teams on delivery plans for AI-assisted engineering, review capacity, and client-ready governance checkpoints.