Sigil: A New Approach to AI Agent Programming

Sigil is setting out to redefine how we program AI agents. It aims to simplify the process by turning conventions into compiler rules, ensuring consistency across the board.

Enforced Order and Naming

The language embraces order. No more subjective argument placements. everything is alphabetical. Parameters, effects, and record fields all follow suit. Types stand out in UpperCamelCase, while everything else, down to file names, sticks to lowerCamelCase.

Nulls and undefined variables are nowhere to be found, replaced by bidirectional type checking and a fat standard library that's still evolving. A notable feature, 'sigil debug', offers replay, stepping, watches, and breakpoints. Meanwhile, 'sigil inspect' gives coding agents a direct line to the compiler, including proof surfaces.

Type Refinements and Constraints

Sigil dives into type theory with solver-backed refinements. It takes traditional numeric types a step further, allowing domain constraints to define types. Here's what that means: you can attach predicates to named types and set proof obligations across function boundaries. The numbers tell a different story when constraints become a language feature.

Consider this example:

t InventoryCount=Int where value≥0

This isn't just syntactic sugar. Underneath, it's powered by Z3, yet the surface remains clean and ordinary. No user-facing proof scripts clutter the experience.

Eliminating Import Chaos

Sigil eliminates import chaos by using rooted references only. This means no more name clashes when pulling in external code. It's a small innovation, but it reduces agent churn significantly. When the model encounters a line, it doesn't have to chase down import statements.

Service dependencies and environment bindings have their specific places in the directory structure. Tests have their syntax and run in parallel, ensuring that every function in a project gets scrutinized. The architecture matters more than the parameter count here, emphasizing structured development.

Constructed with AI Help

Here's a twist: the compiler toolchain wasn't hand-coded. It was generated using tools like Claude Code and Codex. This raises an intriguing question: are we looking at the future of programming? A space where AI not only writes code but also creates the environments we code within?

In the end, Sigil isn't just a language. It's a concept pushing the boundaries of how we think about AI development. By enforcing one way to do things, it's a bold statement against the chaos often seen in programming. But will this strict approach resonate with developers who cherish flexibility?

For those intrigued, explore Sigil atGitHuband see how it's being applied in projects like a toy roguelite.