How YC startups use AI for coding

Coding is where AI stopped autocompleting and started doing the engineering. The clearest proof is Y Combinator's own portfolio: a run of startups from 2023 to today building the AI-native version of writing, reviewing, and shipping software.

Read in order, these nine companies trace one shift, from an assistant that suggests the next line to a team of agents that opens the pull request. What each one automates, the patterns they share, and how to copy the playbook as a small team are below. Company names and batches are public on Y Combinator (see Sources).

The shift: from autocomplete to a team of agents

The old model treats AI as a faster keyboard. It suggests the next line, you accept or reject it, and a human still owns every step of the work: reading the ticket, finding the right file, writing the change, reviewing it, and adding the tests. That version helped, but the human stayed in every loop, so the team's output still scaled with headcount.

The AI-native model hands the agent a whole job, not a line. You describe the task, the agent finds the context across the codebase, writes the change on its own branch, reviews it, and writes the tests, and you approve the result instead of typing it. You can see all three shifts in the companies below: the unit of work moved from a suggestion to a pull request, the agent runs in parallel so a solo founder can supervise several at once, and a review-and-test layer grew up around them so the speed does not cost you correctness.

Nine YC startups building AI coding agents

What they have in common

• The unit of work is a pull request, not a suggestion. The agent takes a described task and produces a reviewable change on its own branch, instead of handing a developer a snippet to paste.
• They run in parallel and isolated. Each agent works in its own branch or workspace, which is what lets one person supervise several at once rather than babysitting a single chat.
• A review-and-test layer grew up alongside the writing agents. The same wave that writes code faster also reviews every PR and backfills the tests, so the speed does not quietly cost you correctness.
• Several sell the layer under the agent, not the agent itself: the fast edit model, the codebase-understanding index, the self-hosted deployment. For coding, owning the context and the guardrails is its own market.

How to copy this as a small team

1. Pick one repetitive job to hand off first, not all of engineering. Test backfilling, small bug fixes, and a first-pass code review are the highest-volume, lowest-judgment work and the easiest place for an agent to earn trust. Module 3 (AI Agents & Automation) frames this as finding the one task worth automating before you scale it.
2. Move from suggestions to whole tasks deliberately. Start by letting an agent open a pull request you review, then run a few in parallel on separate branches once you trust the output, so your time goes to approving changes instead of typing them.
3. Give the agent least access first: read-only on the codebase, then the ability to open a branch and a PR, and never a direct push to main. A human approves the merge, and a CI run plus an AI review gate the change. Treat the agent like a new hire, not a trusted admin.
4. Keep tests and review in the loop, not optional. The point of shipping faster is not skipping the safety net; it is automating it too, so every agent-written change is reviewed and covered before it merges.

Building support this way (AI that resolves, with a human in the loop and least-access by default) is exactly Module 3 (AI Agents & Automation) of AI Operating System for Startups.