Dirac is now sponsored by Unbiased AI.
It was about 8 months ago one afternoon when I felt like a task at hand was challenging my sanity. In normal developer speak, the task was simple and roughly involved adding a new handler with trivial logic, add a DB class, add calls to the handler at multiple places in the codebase. For reasons I would later investigate and find out, it kept growing the context window unexpectedly large, would re-read the same files repeatedly, hit the context limit triggering a crude compaction method, and repeat.
Semantically, you could describe the task very simply and yet I saw it fail twice (note that this was the early 2026 world, a drastically different landscape from late 2026, and I was using Gemini pro). So, what was wrong? A few things. One, several files that had gotten to 1500 ish lines were being pulled in the context where the relevant portion of the code in those files was only a few lines (matter of fact, current OpenAI/Anthropic stacks rely on grep to solve this, coupled with a model better trained to grep all lexical variations of a thing) . Two, there was no infra in place that harness could offer the model a query like ‘find me this class/method from this file’. Three, the only mechanisms to change any code were either rewrite the full file or use search/replace. Four, and this was an interesting failure mode, because the edit used plain search/replace coupled with model shortcomings, edits would often fail or modify 6 locations when the model intended to modify one due to lenient string matches in files like db.py.
I had some background in compilers and syntax tree parsing. So, the obvious low hanging fruits that come to mind were building an AST cache and offering the model capabilities to operate at higher level e.g. find this function, replace this function etc. For any structurally correct codebase, direct AST manipulation generally leads the efficiency-correctness frontier. So, I decided to try it out by adding AST parsers and corresponding harness tooling. I was able to test it the next day, and this alone showed significant jump in capabilities, context use dropped and correctness increased across tasks. Obviously, this is an addictive feeling. So, I kept doing trial-error loops on other high conviction bets (an interesting one is documented here), some of those summarized here.
After a couple of months of hard work with the sole goal of building an agent that I will use in perpetuity, the final product was Dirac. Named after the Paul Dirac, a physicist famously terse with his responses and someone I have a lot of respect for. Also, the Dirac Delta symbol works quite well to convey both precision and 'diffs'.
Fortunately, the hacker news post about Dirac doing well in Terminal Bench was well received. The post alone resulted in a good initial chunk of users, and two job offers. I had to turn those down as I was not looking for employment.
One of the people who reached out to me was David Noël-Romas from Circuit & Chisel, the company behind ATXP and Unbiased. The latter offers the composite LLM called Pareto offering cheaper frontier intelligence.
David kindly offered to sponsor Dirac, and the offer was quite generous. I accepted the offer, and rushed to create Dirac as a corporate entity using Stripe Atlas (yes, Dirac is now a Delaware C Corp). At no point I had planned on this, but a welcome surprise nevertheless.
Rest assured, Dirac will continue to remain provider and model neutral in its functionality so this only affects the harness quality in a positive way. What's more, Dirac users get 1 month free on Unbiased's Personal coding plan 🎉.
So, what’s the plan now? Dirac will continue to be built and improved, it never stopped improving since launch and will continue on this path. Check the homepage to see the ‘book of features’. A big thank you to all the contributors who helped along the way.
The goal was and is to build the best opensource harness that can be used with any model. What “best” means keeps evolving. So will Dirac.