Spec-Driven Agent

Note

Experimental. The Spec-Driven Agent is under active development. Its pipeline, request shape, and defaults may change between releases.

The Spec-Driven Agent (also called the LLM-augmented generator) is the agentic layer that sits on top of BESSER’s deterministic code generators. It bridges the gap between template-based generation and free-form natural-language code generation.

Where a template generator (Django, FastAPI, SQLAlchemy, …) emits a fixed scaffold from your structural model, the Spec-Driven Agent starts from that deterministic scaffold, lets an LLM customise it to satisfy a natural-language request, and then validates the result and repairs what it can. It can add features the templates don’t cover (authentication, JWT, Docker, custom middleware, migrations, tests) or target a stack BESSER has no built-in generator for (Rails, Rust/Axum, Kotlin/Spring, Next.js, …).

The key idea: generation is hybrid, and the model remains the source of truth. Where a BESSER generator fits, the LLM edits its scaffold. For a stack without a built-in generator, it writes the application from stack metadata and the model’s contract.

In this section

When to use it

BESSER offers two ways to turn a model into code. They are complementary, and the agent is built on top of the generators rather than replacing them.

Deterministic generator

Spec-Driven Agent

You provide

A B-UML model, plus the generator’s own options.

A B-UML model and a natural-language description of what to build.

What comes out

One template pass over the model: same input in, same output out.

A deterministic scaffold that an LLM has customised, then validated and repaired — see How it works.

Cost

Free and offline.

Paid LLM calls, or the keyless free tier when the deployment configures one — see Providers and models.

Reach

The ~25 targets listed in Code Generators.

Those targets plus features the templates don’t cover, and stacks with no BESSER generator at all.

Reproducibility

Byte-for-byte reproducible.

Bounded, not reproducible: an LLM authored part of the output.

Reach for a plain generator when the target is one of BESSER’s supported technologies and the templated output is what you want. Reach for the agent when you need something on top of that output, or a stack BESSER does not template.

Limitations

  • Round-trip preservation. Each run regenerates from the model and instructions. Modify mode carries a previous run’s files forward, but hand-written edits made outside BESSER are not re-derivable from the model and are not carried across a from-scratch re-generation. Closing this gap is the most significant planned improvement.

  • Validation has a defined scope. Phase 3 checks syntax, imports, contracts and lint. Authorised sandboxed checks can also boot a supported backend, create records, compile code, and build a frontend. Checks that cannot run are reported as unverified. It does not build a deployment container or prove every business requirement; inspect the findings and test the application in your target environment.

  • Fixes land in generated code. The auto-fix loop repairs the generated files, not the model. A later from-scratch regeneration starts from the model again and will not carry those repairs.

  • Quality depends on the model. Output fidelity tracks the chosen LLM; a weaker or cheaper model produces a thinner result.