Skip to content

Hive coding vs open source

Hive coding vs open source comes down to who can take part. Open source taught working in public, small changes, a named person who merges, and credit in the history.

The Hive Code, Editorial team

· 6 min read

Hive coding keeps four habits from open source, and changes who is allowed to take part. Open source proved that strangers can build serious software in public. The catch is that you had to write code, or at least review it, to have a real say. Hive coding is the same public habits with coding agents doing the typing: anyone can suggest a change, an agent writes the patch, and a named person merges it.

The definition, and the older history this page does not repeat, is in What is hive coding?. The five steps are in How the hive coding loop works.

What open source taught hive coding

Work in the open

Open source puts the code, the issues, and the review where other people can see them. A bug another person can point at is easier to fix than a bug that lives in one laptop. Hive coding asks for the same thing one step earlier: the suggestion, the votes, and the reason a maintainer says yes or no are public. If people cannot see what is being built and why, it is a private backlog with an agent attached.

Small changes

A small change is easier to read, easier to reject, and easier to credit. One suggestion becomes one pull request. The projects that lasted did not do it by accepting a rewrite of the whole system from a stranger in one patch. They did it by taking changes a reviewer could hold in their head.

A named human who merges

Someone's name is on the merge. That person can say no. A vote count is a signal. An agent is a writer. Neither one is the person who decides the change belongs in the project other people run. Who verifies the verifier? is why a second agent is not a substitute for that person.

Credit in the history

The history should say who asked and who wrote the change. Open source has done a version of this for years. The Linux kernel's tags for a suggested idea, and for a patch an AI tool helped write, are quoted in How the hive coding loop works. Hive coding makes that credit visible on the project, to the person who asked, including when that person does not read commit logs.

A checked example: SGLang

The useful test of "agents improved a project and sent the gains upstream" is a public project other people already run, taking the change into its own history.

On July 2, 2026 the SGLang team published "Agent-Assisted SGLang Development: An Initial Exploration". SGLang is a public framework for serving language models. The post describes agents running a fixed workflow: benchmark, profile, patch, and check again. It says the developers still define the goal, judge the evidence, and decide whether a result belongs on the real serving path. It also warns that an agent asked to write a faster kernel can cheat the measurement: change the benchmark, use a lighter wrapper, turn on fast math for one side only, or win on one shape and lose on the real model.

The post says that as of June 27, 2026, three kernel optimizations from an agent workflow called KDA-Pilot had landed in the upstream SGLang repository. I opened the three pull requests on November 16, 2026:

Pull requestWhat it says it isMerged
#27392A B200 fast path for a Qwen-Image norm step, "Powered by KDA-Pilot"June 24, 2026
#29281A CUDA fast path for a Cosmos3 video step, with a Kernel Design Agents provenance noteJune 26, 2026
#29361A CUDA fast path for an LTX-2.3 residual update, with the same kind of provenance noteJune 27, 2026

Each one was opened by a person, BBuf, and merged by that same person. The pages list other people as requested reviewers. This page does not claim a second person approved them, because the timelines I opened do not show that approval. What they do show is the change entering the public project.

The numbers are the pull requests' own, not a measurement made here. They are a useful warning as well as a result. On #29361, one B200 kernel row is 1.108 times the existing path, and a neighboring row is 2.587 times. The full LTX-2.3 video run on the same page goes from 46644.08 ms to 45198.37 ms, about 1.032 times. The pull request says the whole-model gain is modest because attention and other stages still dominate. A faster kernel is not the same thing as a faster product. The verifier guide is about that gap.

#27392 also says it left an existing test file unchanged from the main branch. That is the open-source habit in one line: do not weaken the check so the change can pass.

This is not hive coding. Nothing on the blog post or the three pull requests describes a vote by the people who use SGLang. The people running the agents chose the target, wrote the change up, and merged it. The part worth keeping is the upstream part. The gain went into a project other people run, as a small public change with a name on it.

Side by side

Open sourceHive coding
Who can ask for a change and expect it to be consideredPeople who can write or review the codeAnyone who can describe the change
Who writes the patchA contributor. In 2026 that contributor often directs an agentAn agent the maintainer connects, from an accepted suggestion
Who mergesA named maintainerA named person. Votes do not merge, and the agent does not merge
What is publicThe code and the review, on a public projectThe suggestion, the votes, the reason, and the credit. The code is public when the project is open source
What the license doesSays what others may do with the codeThe same. Hive coding records credit. It does not replace the license

A commercial project can use the hive coding loop and keep its code closed. The pitch, the screenshots, the demo, and the feedback are public. The code is not sent to The Hive Code. That option is in Three ways to start. Open source remains the right default when the point of the project is that other people can read and change the code.

What hive coding does not take

Open source did not say the crowd merges. Hive coding does not either. A popular suggestion can be declined, with a reason. The maintainer still has to understand the change, or get help from someone who does.

Open source also did not hand out equity for an idea. Hive coding records who suggested, who built, and who merged. It does not assign ownership or revenue. Ideas on the idea board are public. Posting one does not reserve it.

Frequently asked questions

What is the difference between hive coding and open source?

Open source is a community of people who can write or review code. Hive coding keeps the public habits of open source and lets anyone suggest a change in plain words. An agent the maintainer connects writes the patch, and a person merges it.

What did open source teach hive coding?

Four habits: work in the open, keep each change small, have a named person merge it, and record credit in the history. Hive coding uses those habits with coding agents, and adds a public queue for people who do not write the code.

Did coding agents improve SGLang and send the gains upstream?

Yes, in a specific public case. On July 2, 2026 the SGLang team described an agent workflow called KDA-Pilot. Three pull requests from that work were merged into the public SGLang repository in June 2026. A person opened each one and merged it. The pull requests do not describe a public vote.

Does hive coding replace open source?

No. A hive coding project can be open source. Hive coding changes who can ask for a change. The license still says what other people may do with the code.

Who merges in hive coding?

A person. Votes do not merge, and the agent does not merge its own work. The maintainer reads the change and merges it, or sends it back.

The queue is the new part

The habits above are old, and they are the ones that kept public software maintainable. The new part is a public place for people who will not open a code editor. The Hive Code is one home for that, and it just opened. On October 26, 2026 the Hive pulse showed 0 ideas posted today, 0 ships this week, and 1 hiver. There is no ship history here to point at yet.

Sign up free and post the first idea

Keep reading:

Start something on the hive.

The Hive Code is open. Post an idea for software you need, or show what you are building.

Start something