Harness Engineering as Categorical Architecture

Bogdan Banu

Home/Archive/arXiv

arXiv · 2605.12239

Harness Engineering as Categorical Architecture

Bogdan Banu

View original source ↗Read source document ↗Export metadata ↓

Abstract

The agent harness, the system layer comprising prompts, tools, memory, and orchestration logic that surrounds the model, has emerged as the central engineering abstraction for LLMbased agents. Yet harness design remains ad hoc, with no formal theory governing composition, preservation of properties under compilation, or systematic comparison across frameworks. We show that the categorical Architecture triple (G, Know, Phi) from the ArchAgents framework provides exactly this formalization. The four pillars of agent externalization (Memory, Skills, Protocols, Harness Engineering) map onto the triple's components: Memory as coalgebraic state, Skills as operad-composed objects, Protocols as syntactic wiring G, and the full Harness as the Architecture itself. Structural guarantees-integrity gates, quality-based escalation, supported convergence checks-are Know-level certificates whose preservation is structural replay: our compiler checks identity and verifier replay, not output-layer correctness or model behavior. We validate this correspondence with a reference implementation featuring compiler functors targeting Swarms, DeerFlow, Ralph, Scion, and LangGraph: the four configuration compilers preserve three named certificate types by identity or replay, and LangGraph preserves the same certificates through its shared per-stage execution path. The LangGraph compiler creates one node per stage using the same per-stage method as the native runtime, providing LangGraph-native observability without reimplementing harness logic. An end-to-end escalation experiment with real LLM agents confirms that the quality-based escalation control path is model-parametric in this two-model, one-task experiment. The result positions categorical architecture as the formal theory behind harness engineering.

Explore related subjects

cs.PL cs.AI math.CT

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Bogdan Banu. 2026-05-12. Harness Engineering as Categorical Architecture. https://arxiv.org/abs/2605.12239

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

From Rocq to Metal: A Pipeline for Formally Verified Microcontroller Firmware

Enforcing invariants in safety-critical firmware is increasingly urgent as generated code becomes widespread, but standard extraction targets for proof assistants require runtimes too large for many embedded devices. We present a pipeline for running formally verified Rocq firmware logic on Cortex-M microcontrollers. The pipeline extracts Gallina to Scheme, compiles it with Encore!, a bare-metal Continuation Passing Style (CPS) bytecode virtual machine, and embeds the result in no_std Rust firmware. We structure applications as pure state-transition functions, so the business logic is proved in Rocq while the event/effect boundary, host callbacks, compiler, and VM remain explicit trusted components. On ST33-class targets with a 50 KB RAM lower bound, Encore! executes Rocq-extracted code end-to-end, stays within the target memory budget on our benchmarks, and validates a transaction-signing application on physical Ledger Flex hardware.

cs.PL↗

arXiv2026

Practical Range Refinement Types with Inference

Refinement types are a static verification technique that aims at increasing the expressivity of traditional type systems while remaining easy and natural to use. While systems based on refinement types have been developed for several mainstream languages, their practical adoption remains limited by their annotation overhead, which is often a more significant burden than when using the "plain" type annotations of languages like Java or Scala. To improve the state of the art, this paper introduces Ranger: a refinement type system designed to keep the annotation overhead small and to seamlessly integrate with imperative-style constructs like variables and loops. As the name suggests, Ranger focuses on integer range types: a particular kind of refinement types that express bounded integer ranges. Such types are widely useful to verify correct index manipulation and in-bounds data accesses, among others. To combine expressiveness and succinctness, Ranger is based on a bidirectional type system, which runs a type inference algorithm to provide the typechecking pass with information useful to reduce the need for user-written auxiliary annotations. Ranger also integrates other forms of lightweight flow-sensitive static analysis techniques that precisely capture the program's behavior without explicit annotations. We implemented Ranger on top of the Licorne experimental programming language. Our experiments show that Ranger's implementation can concisely express and verify a variety of useful properties that fall beyond the capabilities of standard static type systems like those of Java and Scala, and that Ranger compares favorably to other extended type systems, such as the Java Checker Framework and Liquid Java, that can also check properties about ranges.

cs.PL↗

arXiv2026

Djinnlang: Higher-Level Programming by Unambiguous Specification with an LLM in the Compiler

Programmers write formal specifications, and LLMs implement them, proving that each implementation matches its spec. Taken to its extreme, this makes specification languages the new programming languages. We argue that an unambiguity constraint is key: in addition to proving that its implementation satisfies the specification, the LLM must also prove that any other implementation satisfying it must produce the same outputs on the same inputs, i.e. that the relation formed by the constraints is deterministic. This leaves the LLM no leeway on program semantics: as with a conventional compiler, the generated code never needs to be read and can be regenerated from the spec at any time. Under this constraint and with a powerful LLM, the difference between a specification language and a programming language becomes essentially meaningless, and the LLM essentially becomes a part of the compiler toolchain. The arrangement doubles as a strong form of AI control: an untrusted model writes the code, yet its work is tightly checked by a verifier. To demonstrate that our LLM-in-the-compiler paradigm is feasible when supported by our unambiguity constraint, we present Djinnlang, a high-level specification language built for this future. A Djinnlang program consists only of specifications --- the programmer never writes executable code. In place of a traditional compiler, a symbolic translator lowers each spec to Dafny stubs and proof obligations, and a driver harness orchestrates an LLM that fills in implementations and proofs, all checked by the Dafny verifier. We evaluate our language and implementation on multiple examples and we show that it is self-hosting: an LLM can implement the Djinnlang translator from its specification and the reimplementation can verify itself.

cs.PL↗