# Agents Now Start 40% of GitHub PRs. What Indie Devs Should Do

> GitHub's Octoverse 2026 says agents started more than 40% of pull requests in September. What that shift means for bootstrapped developers and small teams.

- Canonical: https://zeroraised.com/blog/github-octoverse-agents-40-percent-prs
- Published: 2026-10-30
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Growth
- Tags: artificial-intelligence, developer-tools, growth

More than 40% of pull requests opened on GitHub in September 2026 were started with an AI agent, up from 17% in March ([GitHub Octoverse 2026](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)). That is 110 million pull requests in a single month, roughly six times the March figure. For a bootstrapped developer, the interesting question is not whether agents are popular. It is what becomes scarce when the first draft of a change is almost free.

**TL;DR**

- More than 40% of pull requests in September 2026 were started with an agent, up from 17% in March.
- Agent-started PRs touched multiple files and merged within two weeks slightly more often than other PRs.
- Repositories with an agent configuration file logged 8x the commits.
- Most developers who choose models use more than one, and Copilot Fusion now routes models per task.
- Treat review capacity, task scoping and tests as the constraints, not code generation.

**Disclosure:** ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 30 Oct 2026; ZeroRaised has no commercial relationship with the companies named.

**An agent-started pull request** is a pull request whose opening change was produced by an AI agent working from a task description rather than typed line by line by a person.

## What exactly did GitHub measure?

The report covers 17.9 billion pushes in the past 12 months, more than 2024 and 2025 combined, plus 34.3 billion commits and 1.3 billion merged pull requests ([GitHub Octoverse 2026](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)). The agent number is a subset of that traffic, not a separate product line.

The jump from 17% to more than 40% is the part worth sitting with. Agent use moved from an experiment to a normal way to begin a change. GitHub reports 55% of Copilot monthly active users using agentic features, up from 35%, and 230% growth in Copilot cloud agent usage ([GitHub Octoverse 2026](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)).

## Are agent pull requests good enough to merge?

On the numbers GitHub published, they look close to parity. About 87% of agent-started PRs change multiple files, against 79% for other PRs. By September, 90% of agent PRs were merged within two weeks, against about 86% for the rest ([GitHub Octoverse 2026](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)).

Read that carefully. A two-week merge window measures whether a change eventually landed, not whether it was good. A team that rubber-stamps a plausible diff will look excellent on this metric. The multi-file stat cuts both ways too: an agent touching four files may be doing real work, or leaving a wider blast radius for a tired reviewer.

The most actionable line in the report is smaller: repositories with an agent configuration file logged 8x the commits ([GitHub Octoverse 2026](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)). That is a correlation, not a causal proof. One plausible reading: teams that wrote down how agents should work in their repo got more out of them than teams that treated the agent as a chat box. Our guide to [reviewing AI-generated code](https://zeroraised.com/blog/how-to-review-ai-generated-code) and the [agent sandbox checklist](https://zeroraised.com/blog/how-to-sandbox-an-ai-agent) cover the practice.

## Why model routing became the default

The report also shows that model choice is fragmenting. Among users who picked a model, 61% used more than one, and 47 distinct models were used in September. No single model had more than 25% of requests. The request mix was 38% general purpose, 37% deep reasoning and 25% fast models, and 63% of paying users used automatic model selection ([GitHub Octoverse 2026](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)).

GitHub has now productised that behaviour. Copilot Fusion is generally available and routes and orchestrates models per task ([GitHub Copilot Fusion changelog](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/)). It offers Standard, High and Max effort levels and two multi-model workflows. In Cascade, an efficient model drafts and a quality check either accepts the work or escalates it to a stronger model. In Critique, one model drafts, another reviews, and the draft is revised ([GitHub Copilot Fusion changelog](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/)).

The economics are the story. Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the API cost per task, about half the tokens and about 40% less time ([GitHub Copilot Fusion changelog](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/)). Billing is per model at standard rates, with a 30% discount for paid plans for a limited time ([GitHub Copilot Fusion changelog](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/)). If you build your own agent loop, this is the pattern to copy: most steps do not need your most expensive model. Our explainer on [what LLM routing is](https://zeroraised.com/blog/what-is-llm-routing) walks through how to do it.

## What should a solo developer change this week?

First, write the agent configuration file. The 8x commit figure is the closest thing to a cheap lever in the report. Put the commands, conventions, test routes and forbidden areas in the repo where the agent reads them, and keep them current.

Second, budget review, not generation. If an agent can open five plausible PRs before lunch, your afternoon is the constraint. Small diffs, one concern per PR and a test that fails before and passes after are worth more than another model upgrade.

Third, measure cost per merged change, not cost per token. Track which tasks needed escalation to a strong model and which were fine on a cheap one. The [LLM cost guide](https://zeroraised.com/blog/how-to-reduce-llm-costs) covers the mechanical side, and the [LLM API comparison](https://zeroraised.com/blog/how-to-choose-an-llm-api) covers the selection side.

Fourth, keep the boring guardrails: lockfiles, pinned tool versions, least privilege for agent tokens and a clean rollback path. Faster generation also means faster mistakes.

## What this means for bootstrapped founders

If agents are writing a large share of the first drafts, the value of a developer shifts from typing to judgment. That has a pricing consequence. Selling "we write code fast" competes with a tool that is getting cheaper every quarter. Selling a narrow, well-scoped outcome, with the review and accountability that go with it, is a better position.

It also changes how you should talk about your product. "Built with AI" is not a differentiator in a world where 40% of pull requests start with an agent. What a customer can check is whether the thing works, whether it is safe and whether you will still be there next month. The [defensible AI product guide](https://zeroraised.com/blog/how-to-build-a-defensible-ai-product) and the [agent versus workflow comparison](https://zeroraised.com/blog/ai-agent-vs-ai-workflow) are the relevant reading.

One caution: shipping more does not fix distribution. If nobody knows the product exists, a faster merge queue just produces more features nobody uses. Keep the [first 100 users work](https://zeroraised.com/blog/how-to-get-your-first-100-users) in the same plan.

## Where ZeroRaised fits

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have something people can try, [submit your product](https://zeroraised.com/submit); the [pricing page](https://zeroraised.com/pricing) explains the current options. A launch page is a plain place to say what the product does, who made it and what it costs, which is exactly what a buyer wants after the demo.

For more analysis like this, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter). No promises of traffic or results, just the useful parts.

## Sources

- [GitHub Octoverse 2026: commits and pushes reach new highs as developers build with agents](https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/)
- [GitHub changelog: route and orchestrate models with GitHub Copilot Fusion](https://github.blog/changelog/2026-10-28-route-and-orchestrate-models-with-github-copilot-fusion/)

## Frequently asked questions

### What share of GitHub pull requests are started by agents?

GitHub's Octoverse 2026 report says more than 40% of pull requests in September 2026 were started with an agent, up from 17% in March. That was 110 million pull requests in September, about six times the March figure.

### Are agent-written pull requests merged less often?

The report suggests the opposite at the two-week mark. It says 90% of agent-started PRs were merged within two weeks in September, compared with about 86% of other PRs. That measures merge speed, not review depth or long-term code quality.

### What is Copilot Fusion?

Copilot Fusion is GitHub's generally available feature for routing and orchestrating models per task. It offers Standard, High and Max effort levels and two multi-model workflows, Cascade and Critique.

### Does this matter if I am a solo founder?

Yes, because the bottleneck moves. If an agent can open a plausible pull request in minutes, your output is limited by how quickly you can define a task and review the result. That favours small teams with tight scope and good tests.