About ARCHON
Inspiration
I started building ARCHON because I noticed a gap in how we use AI for software development.
AI can generate code surprisingly well. But when a real project breaks because of a dependency update, a deprecated API or an unexpected runtime error, getting a suggested fix is only half the job. Someone still has to investigate the error, find the right documentation, apply the patch, run the tests and figure out what broke next.
I wanted to build something that could handle that entire process instead of stopping after generating code.
That became ARCHON: an autonomous software repair agent that investigates bugs, researches relevant documentation, proposes fixes and tests those fixes in isolated execution environments before presenting the results to the developer.
The idea was simple: an AI agent should not claim that it fixed a bug just because it generated a plausible patch. It should have evidence that the patch works.
How we built it
ARCHON is built around a repair loop that combines AI reasoning, live research and actual code execution.
We use NVIDIA Nemotron models through Nebius Token Factory, with different models handling different responsibilities. Nemotron 3 Ultra handles the more demanding diagnosis and patch generation, while Nemotron 3 Super helps supervise the process, research errors and review candidate fixes. Nemotron 3 Nano helps compact and process test output.
For research, we integrated the Tavily Search API directly into the workflow. Rather than relying entirely on a model's existing knowledge, ARCHON searches for relevant documentation, changelogs and issue discussions associated with the failure.
The most important part is the execution environment. Using Nebius Token Factory Sandboxes, ARCHON can reproduce failures, apply candidate patches and run tests in isolated environments. Each candidate gets its own sandbox fork, so a failed attempt does not contaminate another attempt.
The application itself consists of a Next.js cockpit and a FastAPI backend. The interface streams the agent's progress, terminal output, research activity and code diffs, allowing developers to see what happened instead of receiving a mysterious final answer.
ARCHON currently supports two workflows: automated bug repair and migration of applications from closed LLM providers to Nebius Token Factory.
Challenges we faced
One of the hardest parts was learning that a patch can look correct without actually being correct.
Models sometimes produce malformed edits, miss important source details or generate changes that do not fix the underlying failure. We had to build validation into the workflow rather than trust the model's response.
We also encountered a subtle problem with AI-provider migrations. When a repository uses mocked clients in its tests, the test suite can pass even when the actual client implementation was never migrated. We addressed this by introducing structural checks that inspect the source code for leftover closed-provider client constructions and model references.
Working with sandbox execution brought its own challenges, including environment setup, test-command handling, SDK differences and interpreting test results reliably. We gradually made the system more defensive, added test coverage and improved how it selects candidate patches.
Not every mission succeeds. For example, our more complicated Anthropic migration fixture remained unresolved after repeated attempts within the configured budget. Rather than label an incomplete migration as successful, ARCHON reports the failure. That distinction matters when the tool is supposed to be trusted with real code.
What we learned
The biggest lesson was that building an agent is not just about choosing a powerful model. The surrounding system matters just as much.
Reliable execution, good feedback loops, isolated experiments, source-grounded research and honest verification are what turn a model into a useful engineering tool.
We also learned that passing tests are meaningful only when they actually test the behaviour we care about. Verification needs to be designed into the system, not added as an afterthought.
In our initial SWE-bench Verified experiments on the psf/requests repository, ARCHON resolved six out of six selected instances in Nebius Token Factory Sandboxes. These are results from our current test set, not a claim of solving the full SWE-bench benchmark.
What's next
Our goal is to make ARCHON a dependable engineering assistant that developers can use to investigate broken builds, repair existing code and migrate AI applications with greater confidence.
We want every repair to come with a clear explanation, a reviewable diff, an execution transcript and test results that developers can inspect before accepting the change.
ARCHON is built around a principle we want to keep simple: don't just generate the fix. Run it, test it and show your work.
Log in or sign up for Devpost to join the conversation.