Anyone who runs a public repository knows the effect: the more visible a project becomes, the faster the list of open items grows. This is precisely where AutoCraft comes in, a new open-source tool intended to process issues and pull requests with the help of AI. It emerged from the environment of ArtCraft, a project that attracted attention with AI-supported recreations of well-known Adobe applications – and as a result of which a growing backlog accumulated on GitHub.
The starting point is therefore less a technology experiment than a capacity problem. A backlog, i.e. the stack of as-yet unprocessed tasks, bug reports and improvement suggestions, cannot be resolved at will by more working hours. The idea behind AutoCraft is accordingly: part of this work should be taken over by software agents that independently pick up tickets and generate change proposals.
What an agent in the repository can actually do
An AI agent is understood to be a program that not only has a language model generate text, but gives it tools – such as file access, terminal commands or the API of a platform like GitHub. The agent therefore reads a task description, gains context in the source code and uses it to generate a concrete change proposal.
In a repository context, the benefit typically lies not with the difficult architectural questions, but with the drudgery that otherwise remains undone:
- triaging incomplete bug reports, i.e. sorting, labelling and adding follow-up questions
- recognising obvious duplicates and items that have already been fixed
- preparing small, clearly scoped corrections as a pull request
- updating documentation when interfaces have changed
- updating dependencies and describing the impact
The decisive point: a pull request is a proposal, not a publication. Handover to a human is already built into the workflow. This makes repository agents a comparatively low-risk entry point for automation – provided the review actually takes place.
Why open source in particular produces this pattern
Open-source projects are an almost ideal breeding ground for such tools. The code is public anyway, the tasks are structured in a ticket system, and the contribution process via pull requests is standardised. At the same time, the pressure is high: the number of reports scales with the number of users, the number of maintainers generally does not.
The fact that a project which itself relies heavily on generative methods is now also supporting its project organisation with AI is consistent in this respect. But it also shows a dynamic that companies should be aware of: if creating software becomes cheaper, the bottleneck shifts to review, maintenance and quality assurance. More output creates more maintenance load – and thus the need to automate that load as well.
The open questions
Reliable statements about the quality of the results cannot be made at this point in time. From the practice of comparable approaches, however, some topics are foreseeable that need to be clarified with every introduction:
- Review capacity: Automatically generated pull requests shift effort, they do not eliminate it. Anyone who generates ten proposals a day but can only review three has merely redistributed their backlog.
- Permissions: An agent needs access tokens. These should be scoped as narrowly as possible – write rights on branches, not on the main branch, no approval of releases.
- Prompt injection: Issue texts come from third parties. If an agent interprets them as instructions, its behaviour can be influenced from outside. Inputs from public sources must fundamentally be treated as untrusted.
- Costs and transparency: Model calls cause ongoing costs. These should be measurable per repository, otherwise there is no basis for deciding whether the use pays off.
- Traceability: Machine-generated contributions should be recognisable as such so that a project's history remains interpretable years later.
A realistic start
For companies that want to test the approach, a step-by-step procedure has proven effective. Initially, tasks without production risk are suitable: documentation, test coverage, formatting, dependency updates. Only when the hit rate has been stable over several weeks does the step to functional changes pay off.
Preparation in the repository itself is also important. Agents benefit from the same things as new team members: clearly described tickets, a readable project structure, documented conventions and automated tests that stop a faulty change early. Where these foundations are missing, automation mainly creates additional work.
And finally, a metric is needed. What makes sense is not the number of pull requests generated, but the proportion that reaches the main branch without significant rework. Only this rate shows whether an agent relieves or occupies.
Assessment
Tools like AutoCraft mark the point at which AI in the development process moves from code completion to process participation: the agent no longer works in the editor, but in the ticket system. For web and software projects, this means that maintenance effort will in future be determined more by review capacity than by development time – and that clean tickets, tests and permission concepts become the prerequisite for any automation gain.
Anyone who invests in these foundations today can later switch such agents on in a controlled way instead of having to catch them after the fact.
Sources
- After the Adobe clones: ArtCraft unleashes AI on GitHub backlogs — heise developer News