Governed execution for trustworthy AI agents

ADL: the language

Today, agent systems live in scattered prompts and glue code — impossible to inspect, version, or audit. The Agent Design Language (ADL) defines them explicitly: agents, tools, permissions, workflows, and validation rules, declared in one specification a human can read and a runtime can enforce.

ADL Runtime: the control plane

The runtime executes what the language declares. It enforces policy, gates tool use, and records every prompt, call, and decision as a replayable record — deterministic control over non-deterministic models. This is the control plane for enterprise agent systems.

CodeFriend

Our first product, and the platform's proving ground: AI-native software engineering built on ADL. Agent teams deliver code through the Cognitive SDLC — every change bound to an issue, validated, and reviewed from four independent perspectives: correctness, security, adversarial, and policy compliance. The runtime keeps the evidence; your engineers keep authority. In development now, with early access through the design-partner program. Learn about CodeFriend - launching soon!

Recursive self-improvement, bounded

Our Gödel agents run the loop most labs only theorize about: observe their own performance, hypothesize improvements, test them, and adopt what survives. The Gödel–Hadamard–Bayes algorithm makes self-improvement a governed, replayable process — every revision recorded in the reasoning graph, every adoption inside the same gated runtime as everything else. Self-improvement as an engineering discipline, not an article of faith.

Reasoning graphs & loops

Our agents record what they believed and why — claims, evidence, criticisms, hypotheses, and revisions in a graph you can interrogate for contradictions, unsupported conclusions, and unanswered objections. And the graph isn't passive: when it shows tension, the loop runtime generates hypotheses, runs discriminating tests, and revises beliefs. The scientific method, running as software.

The Freedom Gate

Alignment isn't only a model property — it's an architecture property. Every consequential action an agent proposes passes through the Freedom Gate: a policy-and-evidence checkpoint that can allow, refuse, or route to a human. Refusals are recorded with reasons, approvals leave evidence, and operator authority is enforced by the runtime — not requested politely in a system prompt.

A Mathematical Theory of General Intelligence

We propose a mathematical framework for measuring general intelligence across biological and artificial systems. Intelligence is treated not as an abstract property, but as a measurable quantity: the efficiency with which a system discovers and exploits compact, task-relevant representations while operating within explicit resource constraints. Building on state-space compression (SSC), we develop a formal model of tasks, admissible representations, and resource-aware optimization, leading to Cognitive Compression Cost (CCC)—the minimum achievable cost of solving a task under a declared modeling policy.

Published June 2026 · DOI: 10.5281/zenodo.20738007 ·

The Cognitive Software Development Lifecycle

Every SDLC ever written assumes humans write the code. C-SDLC is the first lifecycle designed for AI-centric production: it shows how parallel agent teams can approach the Amdahl's Law limit of development speed — and what breaks, and what becomes possible, when the marginal cost of software falls toward zero. For enterprises, it's a blueprint for multiplying SDLC throughput without surrendering quality, review, or control. CodeFriend is where this lifecycle becomes product.

Preprint in review

Universal Tool Schema (UTS)

Tool calls are where agents touch the real world — and every provider defines them differently. UTS is a proposed open standard for provider-neutral tool interfaces: the same tool runs on any runtime, every call can be recorded and replayed, and side effects are declared and governed rather than improvised. It's the interoperability layer a governed control plane needs underneath it.
Read about UTS 1.1 ->