Ask HN: Is anyone building real software with AI agents?

3 points by variety8675 18 days ago | 10 comments

I've noticed a pattern where the people who are talking about how impressive their agentic workflows are, always seem to use these workflows to build more AI tooling.

Has anyone seen a project built by an agentic workflow that could stand on its own?

aurareturn 18 days ago |

Yes. We rewrote our entire internal tools backend and frontend with Claude in the last 6 months. This was a tool with about 15 years of development by humans.

We did not rewrite our customer app but all maintenance and new feature code is 99% AI agent written and has been for the last 6 months.

You're not going to see us write a blog post about it though. We prefer not to tell our customers that the software they use is mostly AI written.

jonahbenton 18 days ago | |

With what process, spec driven? Having AI produce all the code is pretty common. An agentic workflow operating at a higher level of abstraction especially for brown field is not.

aurareturn 18 days ago | | |

The most important thing to this is people who have domain knowledge. The same leaders who built the original also built the new AI-written version. We know how it needs to work and we know when the AI goes off rails so we need to pull it back.

There was a front page post on HN here a few weeks ago where it said something like the moat is domain knowledge. It was true for us when we rebuilt the internal tool from scratch.

variety8675 18 days ago | |

I like the there is no benefit to talking about it for most companies angle, I didn't consider that

zhoBEENG 18 days ago | | |

It’s also worth considering that a lot of this software will never be marketed, compared to in the past. For example, since it is now trivial to build a tool, why sell it? It is now economical to build a lot tools in-house.

zhoBEENG 18 days ago |

I used claude for a small project last weekend that will 100% create significant value in my workplace. It’s a non-tech setting. Going to be vague so as not to give out too much personal information.

Basically it allowed me to easily spin up a highly usable front-end as well as scrape together a large dataset to back it. The dataset is industry specific knowledge that does not exist in any public, single source. Part of this dataset creation was generating geolocation polygons based on legal documents that verbally defined physical territory. Absolute precision is not necessary for our purposes, but as far as we have confirmed so far the agent basically one-shotted the data with negligible errors and sufficient precision.

Humans were involved in the process, but there is zero chance this project would have happened without agents. The dataset alone would have taken weeks of full time effort. The front end would have been easier, but still time consuming.

I don’t really understand how people aren’t seeing the value in this.

variety8675 18 days ago | |

I see value in using the models for sure, what I'm questioning is all of the examples of basically trying to one shot a Jira ticket to code

zhoBEENG 18 days ago | | |

Yeah, I did not use Jira or other project management software. There was no need as the project could be implemented very fast with essentially one contributor.

However agents (not just LLMs) were definitely involved.

It’s possible I misunderstood your question.

ken-jo 18 days ago |

It seems that companies are still hesitant to market themselves as having built products with AI agents, so they find it difficult to make such projects public. At the moment, AI products appear to be emerging mostly at the level of personal projects or simple websites. I think it won’t be long before we start seeing independent projects being prepared.

xxwink 17 days ago |

Personally yes. And I'm sure others have too. The thing is: Large enterprises cracking the agentic swarm isn't giving away their trade secrets. Also not giving away too much info myself here, because it might be flagged as promotional.