Overview & result
Augustus is a set of agent skills for deciding where a decision model belongs in a system, then building, evaluating and improving that placement against real outcomes. Its README names TypeSafe Jev (Choice, Score, Noul) as the default hosted example, alongside classical classifiers, rankers and other model families, and says the project is independent of TypeSafe.
From the original post
Agent skills for designing, training, evaluating and improving application-specific decision systems. Primitive/model selection, data assembly, export/reload and bounded hill climbing. TypeSafe Jev is the default hosted exemplar; independent of TypeSafe.
How Jev fits in the loop
- You ask your agent to audit, design or improve a workflow using Augustus.
- The skill follows one working model: evidence, then a bounded judgment, then an explicit policy, a checked action and an observed outcome.
- Advice requests return a short design card and a falsifier; build requests return working adapters, policy and evals.
- Improvement requests set up a bounded incumbent-versus-challenger loop with confirmation data kept untouched.
How to reproduce
- Inspect the original showcase and source material at https://github.com/24601/Augustus
- Verify the bounded prompt and input candidate schema configured for Jev
- Benchmark decision latency and classification accuracy against baseline models
Why this build matters
It treats Jev as one option among several model families and says outright that the best result is sometimes a formula, a checklist or no new model.
Limitations
- Augustus is independent of TypeSafe; the README leaves API contracts to the official TypeSafe skill and provider docs.
- It is a skill that guides an agent, so results depend on the agent following it.
- The README reports no measured results from using it.