State of the art
A living map of the work that structured intent draws on, organized by the parts of an intent. Last updated 7 November 2026.
For each new advance as it happens, see the news. Suggest an addition by opening an issue on GitHub, with a link to a primary source.
How to read this map
Structured intent combines several properties in one artifact (see the manifesto). Each row below is one property. Each entry names work that already provides part of it. The last column says what is still missing.
| Part of an intent | Who already does part of it | What is still missing |
|---|---|---|
| Contract (requires, ensures, never) | Design by contract (Eiffel); formal specification languages such as Dafny, JML and TLA+; LLM spec generators such as SpecGen and AutoSpec | Contracts tied to priorities, origin and evidence in one portable file |
| Spec as source of truth | GitHub Spec Kit (Sept 2025), AWS Kiro (July 2025), Tessl; the spec-as-source model (Jan 2026) | Specs mostly live as Markdown prose; little machine-checkable structure; mostly new code, not existing systems |
| Priorities | Requirements engineering (prioritization, rationale); the OpenAI Model Spec for model behavior | Priorities as a first-class part of a software spec that bindings must respect |
| Origin (intended, accidental, bug) | Requirements traceability; decision records (ADRs) | Clause-level rulings recorded during lifting from legacy code |
| Lifting from existing code | Syzygy and SPECTRA (spec mining during translation); DARPA TRACTOR and Intel IDEAS (C to Rust) | The recovered spec is scaffolding for one translation and is then discarded |
| Merging several systems | Software product-line engineering (feature models) | AI-assisted merging of behavior from several codebases into one intent |
| Multiple bindings | Model-Driven Architecture (OMG, 2001); interface definition languages such as Protobuf, OpenAPI and Smithy; Dafny’s multiple compilation targets | MDA never became mainstream; IDLs cover data shapes, not behavior |
| Evidence per binding | Verification tools (Kani, Verus, Prusti for Rust; CrossHair for Python); assurance cases in safety-critical engineering | A standard way to record, per clause and per binding, how a guarantee was established and under what assumptions |
| Equivalence checking | VERT (bounded equivalence via Wasm); Heimdall (formal equivalence for eBPF) | Checking each implementation against an intent, rather than against each other |
| Spec quality | Mutation-based completeness in SpecGen and KBSpec; Spec-Harness, which shows verifier acceptance overstates quality | Quality scores carried inside the spec itself |
| Intent formalization | Lahiri, Microsoft Research (March 2026): intent formalization as the key challenge for reliable AI coding | Priorities, origins and multiple bindings beyond formal specs for single programs |
| Intent engineering | Practitioners including Daniel Miessler, Nate Jones and Vincent Koc | Mostly prompting and organizational alignment; no formal artifact |
| The term “structured intent” | Sean Grove, OpenAI (July 2025): “the valuable artifact is structured intent” | A defined discipline and notation around the term |
Changelog
- 7 November 2026: first version.