# The state of agentic engineering

Agents now start more than 40% of pull requests on GitHub. Output is up, delivery is not, and review is the new bottleneck. What the data says in November 2026, and what working teams do.

- Published: 2026-11-14
- Running time: about 5 minutes (animated)
- Watch: https://aikaara.com/explainers/state-of-agentic-engineering
- Publisher: aikaara (https://aikaara.com)

## Opening

### AI writes the code now. Who checks it?

A short film for CEOs and CTOs on where agentic engineering stands in November 2026: what the data shows, where engineering time goes now, and what the teams getting results do differently.

Started by an agent
Started by a person
40%+

of pull requests on GitHub started with an agent’s first draft in September 2026.

In March 2026 it was 17%. The share more than doubled in six months.
Writing code has become cheap. Checking it has not. That gap is the subject of this film.

Sources: [GitHub Octoverse 2026 (vendor data)](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/); [DX Q3 2026, 500+ companies (vendor data)](https://getdx.com/uploads/state-of-ai-impact-in-engineering-q3-report.pdf)

## Chapter 1: Agents arrived, fast

### Agents now start a large share of the work

Share of pull requests on GitHub that started with an agent’s first draft.

Pull requests started by an agent

17%
40%+
GitHub counts a pull request when an agent wrote the first draft. A person may still have changed it before merging.
GitHub says its detection is a floor, so the real share may be higher.

Sources: [GitHub Octoverse 2026 (vendor data)](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)

### Most developers now use agents at work

The same survey question, asked a year apart.

Developers using AI agents at work

31%
59%
Developers using AI agents at work went from 31% in 2025 to 59% in April 2026.
This is a pulse survey of 1,100 developers, so treat the exact figure with care. The direction matches GitHub’s telemetry.

Sources: [Stack Overflow pulse survey, May 2026, n=1,100](https://stackoverflow.blog/2026/05/27/agents-on-a-leash-agentic-ai-remains-mostly-monitored-at-work)

### Used everywhere. Trusted narrowly.

What developers say about trusting AI output.

Trust AI when they can easily check the output
48%
Trust AI with important work decisions
6.6%
949 of 14,304 answers

Developers trust AI where they can check it. The question for a company is who does the checking, and how fast.

Sources: [Stack Overflow Developer Survey 2026, n=14,304](https://survey.stackoverflow.co/2026/ai/data)

## Chapter 2: More code, not faster delivery

### Developers felt faster. They were slower.

A randomised trial in 2025, with experienced developers working on code they knew well.

Effect of AI on task time

20% faster
19% slower
Experienced developers took 19% longer with AI. Afterwards they believed they had been 20% faster.
The tools were early-2025 tools. The lesson that lasts is that self-reported speed is a poor guide.

Sources: [METR, July 2025 (randomised trial, 16 developers)](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)

### The trial could not be repeated

In 2026 METR tried to run the study again.

2025: the trial ran

16 experienced developers, tasks assigned at random to AI or no AI.

2026: it could not run cleanly

57 developers. Many declined or held tasks back because they would not work without AI.

In 2026 METR could not repeat the trial cleanly. Developers would not work without AI.
Many declined to join or held tasks back. METR calls the new result only very weak evidence.
Anyone who quotes a clean 2026 speed-up from this work is misreading it.

The most careful measurement of AI speed now has no clean control group.

Sources: [METR, Feb 2026](https://metr.org/blog/2026-02-24-uplift-update/)

### Output per developer is up sharply

Vendor telemetry from more than 500 companies.

75.5%

rise in complexity-weighted pull requests per developer in a year.

DX Q3 2026, vendor data
DX weights each pull request by how complex it is, so this is more than a raw count.
More output is the start of the story. The next chapter is what happens to it.

Sources: [DX Q3 2026, 500+ companies (vendor data)](https://getdx.com/uploads/state-of-ai-impact-in-engineering-q3-report.pdf)

## Chapter 3: The review queue

### Code waits longer to be merged

The strongest study so far: 655 firms, before and after they adopted coding agents.

56%

longer from pull request to merge, a year after adopting coding agents.

Harvard working paper, Nov 2026
A year after adopting coding agents, time from pull request to merge rose 56%.
This is a working paper, not yet peer reviewed. It is the largest study of its kind.

Sources: [Chen & Stratton, Harvard working paper, Nov 2026, 655 firms](https://fion.ac/jellyfish.pdf)

### The more AI code, the longer the wait

Median wait for the first human review, by share of code written with AI.

Wait for a first review

Teams with 20% to 40% AI code
40 min
Teams with 80%+ AI code
82 min
Teams writing the most AI code wait twice as long for a first review. 82 minutes versus 40.

Reviewers are now the scarce resource.

Sources: [DX Q3 2026, 500+ companies (vendor data)](https://getdx.com/uploads/state-of-ai-impact-in-engineering-q3-report.pdf)

### Some teams clear the queue by skipping it

One way to cut the wait is to stop reviewing.

76%

more pull requests merged with no review at all.

Faros AI, Sep 2026, vendor data
Pull requests merged with no review are up 76%.
No review, human or agent, before the code went in.

Skipped review is where fast output turns into risk.

Sources: [Faros AI, Sep 2026, 22,000 developers (vendor data)](https://www.faros.ai/blog/ai-speed-trap-takeaways)

## Chapter 4: Quality and risk

### Faster and less stable at the same time

Google’s DORA research, from about 5,000 professionals.

Up

Delivery throughput.

Up

Delivery instability.

A direction, not a percentage
AI raised delivery throughput.
It also raised delivery instability.
DORA reports this as a direction, not a percentage. We do not put a number on it.

Speed without stability is not a gain a customer feels.

Sources: [DORA 2025, ~5,000 professionals](https://services.google.com/fh/files/misc/2025_state_of_ai_assisted_software_development.pdf)

### The speed fades. The warnings stay.

Open-source projects before and after adopting Cursor, against matched projects that did not.

2

months of faster work, then the gain faded.

Open-source projects
30%

more static-analysis warnings, and they stayed.

Peer reviewed
After adopting Cursor, open-source projects sped up for about two months. Static-analysis warnings stayed about 30% higher.

Sources: [Carnegie Mellon, MSR 2026, 806 repos](https://arxiv.org/abs/2511.04427)

### Security has not caught up

A benchmark of code written by AI models, with no review or security tooling.

AI models’ code that passed security checks
56%
Unchanged since 2025
AI models’ code passed security checks 56% of the time. That has not improved since 2025.
This tests raw model output, before any review or CI. Closing that gap is the job of your pipeline.

Sources: [Veracode 2026 benchmark (vendor data)](https://www.veracode.com/news/llms-are-getting-smarter-but-not-safer-veracode-2026-genai-code-security-report-finds-ai-generated-code-security-has-stalled-at-56%25-pass-rate/)

## Chapter 5: The bill before the proof

### AI coding is now a $34 billion line

An estimate of what enterprises spend on AI coding each year.

$34B

a year on AI coding, an estimate.

Menlo Ventures, an Anthropic investor
7.7x

growth in a year.

Same estimate
Enterprises spend an estimated $34 billion a year on AI coding. Up 7.7x in a year.
This is a venture firm’s model, not an audited figure. The firm invests in Anthropic.

Sources: [Menlo Ventures estimate, Nov 2026 (VC, Anthropic investor)](https://menlovc.com/wp-content/uploads/2026/11/menlo_ventures_enterprise_ai_report-2026-110326.pdf)

### Cost is now usage, not seats

Agents bill by what they consume.

$150

to $250 a month for a typical developer on Claude Code.

Anthropic’s own figure
A typical developer on Claude Code costs $150 to $250 a month.
Heavy users and teams of agents cost more. The bill moves with how the tools are used.

Ask for cost per merged pull request, not cost per seat.

Sources: [Anthropic Claude Code docs (vendor data)](https://code.claude.com/docs/en/costs)

## Chapter 6: What working teams do

### AI is an amplifier

DORA’s summary of what it found.

Seven capabilities that make AI pay off

1
A clear AI stance
2
Healthy data
3
AI access to internal knowledge
4
Strong version control
5
Small batches
6
A focus on users
7
Good internal platforms
A leadership decision
An engineering habit
AI magnifies strong engineering systems and weak ones.
Five of the seven are long-standing engineering habits. Two are leadership decisions: stating a stance, and giving the tools your company’s knowledge.

Sources: [DORA 2025, ~5,000 professionals](https://services.google.com/fh/files/misc/2025_state_of_ai_assisted_software_development.pdf)

### Stripe: agents write, humans review

One of the best-documented cases.

1,300+

pull requests a week with no human-written code.

A human reviews every one
Stripe merges over 1,300 pull requests a week with no human-written code. A human reviews every one.
Each agent works in an isolated machine, with short scoped rules and a small set of tools.
Lint runs before push, and full CI runs before a person looks at it.

Sources: [Stripe engineering blog, Feb 2026](https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents-part-2)

### Uber: watch the cost of each change

Another documented case, focused on spend.

Engineers are alerted as they use their expected spend

First alert
50%
Second alert
80%
Final alert
100%
Uber routes its agents’ tools through one gateway.
Cheaper models do the routine sub-tasks by default.
Uber tracks cost per merged pull request.

Both companies keep agents inside the same tests, CI and review gates as people.

Sources: [Uber engineering blog, Aug 2026](https://www.uber.com/blog/efficient-software-factory/); [Stripe engineering blog, Feb 2026](https://stripe.dev/blog/minions-stripes-one-shot-end-to-end-coding-agents-part-2)

### Six questions to ask your CTO

How long does a pull request wait for its first review?

If you do not know, you cannot see the bottleneck.

How many pull requests merged last month with no review?

This is the number that turns speed into incidents.

What does one merged pull request cost us in AI spend?

Seat counts hide the real bill.

Do agents run with the same tests, CI and permissions as people?

Agents should not get a side door.

What did delivery and quality look like before we rolled out agents?

Without a baseline, nobody can say whether it worked.

Who owns review capacity as a resource?

Review is now the scarce work. Someone has to plan it.

### The short version

Agents now start more than 40% of pull requests on GitHub.
Output per developer is up. Time to merge is up too.
Review is the new bottleneck, and some teams skip it.
Spend is measured. Return mostly is not.
The teams getting results build the checking system first.

Data as of 14 November 2026. Vendor and investor data is labelled. Every figure is linked under Sources.
