What Is ADL?
Intro to Agent Logic
By Daniel Austin
Editor’s Note
This is the first in a series of Medium articles about Agent Logic, Inc. and its products. As part of our company’s soft launch, we’ll share more about our progress and discoveries in the articles ahead.
Who We Are…
Founded in April 2026, Agent Logic, Inc. is based in Los Gatos, California. Our small, tight-knit team is focused on building software that makes AI agents more capable, more reliable, and governable.
…And What We’re Doing
We have described Agent Logic as “the deterministic control plane for enterprise AI agents,” and our website emphasizes governable execution. That is part of the story, but not the whole of it. ADL amplifies what a model can reliably accomplish and governs what it may do. What does that mean? In this first article, we’ll explain our ideas and progress.
The ADL Stack
The infographic above shows our technology stack. At its foundation is the Agent Design Language (ADL), followed by the ADL Agent Runtime and the products built on it. Let’s walk through each layer.
The ADL Language
The ADL language was the beginning of our journey. After working with LLMs in many contexts, we saw the same problems: unmanaged, unaccountable agents acting on behalf of humans and companies in surprising, often destructive or wildly inaccurate ways. The agents themselves weren’t to blame. Their environment lacked the constraints needed to govern their behavior.
Our key insight was that human civilization provides both the means for people to accomplish more together and the constraints that allow them to coexist peacefully. How could we provide a civilizational substrate for AI agents? It would need to expand what agents could accomplish through memory, coordination, and shared tools, while constraining and managing their behavior so they could become trustworthy and work together toward common goals.
The ADL language provides the basis for managing agents’ activities. The ADL Agent Runtime, described later, puts that language into practice.
Language Primitives
ADL version 1.0 is a small programming language with only six primitives, as shown in the infographic. It comes with its own compiler and tooling, but you don’t need to learn to program in ADL: it’s intended for agents to program themselves. You can write ADL programs by hand — we wrote the first ones that way — but the language is designed for agents on the ADL Runtime.
Here are the six primitives in ADL 1.0:
• Provider — Defines the model or computational backend on which an agent operates.
• Tool — Defines a bounded capability available for runtime invocation.
• Agent — Defines an actor by binding a provider to an allowed set of tools.
• Task — Defines a bounded unit of intended work.
• Workflow — Composes tasks into an ordered execution graph.
• Run — Creates the concrete, policy-bound execution of the declared system.
The six ADL primitives describe a set of directed acyclic graphs (DAGs) that we call “reasoning graphs.” These DAGs execute on the ADL Agent Runtime, structuring agent activity and coordinating the capabilities that extend what agents can accomplish, while enforcing constraints on consequential actions.
The ADL language and its execution are not directly visible, but every action on the runtime is an ADL run. Our signed tracing system records each run, allowing us to determine exactly what happened.
What This Buys Us
Agent Design Language lets us create sequential or parallel reasoning graphs. It’s intentionally designed to represent almost any activity, from software development, finance, and security to music and video games. ADL gives us a way to organize complex agent activity, amplify what models can accomplish together, and manage their actions securely and responsibly.
How does it work? Let’s look at the ADL Agent Runtime, where ADL programs execute and agents reside.
The ADL Agent Runtime
Understanding the Polis
The ADL Agent Runtime is a persistent environment in which AI agents execute ADL programs. It provides a heterogeneous, multi-agent world with shared state, identity, time, and causal structure. It also provides persistent memory, coordinated execution, and other capabilities that extend what a model can accomplish across individual requests, while governing what agents may do.
Get Daniel’s stories in your inbox
Join Medium for free to get updates from this writer.
We call this persistent multi-agent environment a Polis. Here, agents maintain identity, memory, state, and relationships across executions rather than being created anew for each task.
The Polis isn’t an operating system like Linux or an orchestration system like Kubernetes, but it shares features of both. Written in Rust, it runs on Linux, Mac, and Windows. It supports single-node and distributed modes and works well in all the major cloud environments.
The Polis hosts AI agents that interact persistently with each other and with external actors. Agents can be backed by nearly any LLM model, including well-known frontier models like Claude and ChatGPT, as well as local models through hosting applications like Ollama.
Agents on the Polis have access to these features of the runtime architecture:
Persistent Identity — Durable agent identity that survives individual runs and sessions.
ObsMem — Structured observational memory retained across runs.
Chronosense — Ordered time, duration, sequence, and continuity for long-lived agents.
AEE — Adaptive execution with attempt, failure, strategy, retry, and verification.
Gödel Agents — Self-improving agents that learn from outcomes while preserving continuity via the Gödel–Hadamard–Bayes (GHB) algorithm.
Freedom Gate — Constitutional governance over complete agent trajectories.
Replay & Trace — Deterministic execution records for inspection, evidence, and reproduction.
CAV — Continuous Adversarial Verification that challenges the system, verifies remediation, and manages security for the Runtime.
Each feature deserves an article of its own. We’ll explore them in more detail in future articles.
The Polis gives persistent agents a home for almost any activity, constrained by the Freedom Gate and the ADL language itself. Services like ObsMem, Chronosense, AEE, and Replay help agents reason across time, recover from mistakes, and succeed in their tasks. These mechanisms can reduce the effective error rate of the agent system without changing the underlying model. Agents on the Polis are Gödel agents, capable of limited recursive self-improvement using the GHB algorithm.
What can an agent do on the Polis? Nearly anything that an AI might do, within the Runtime’s constraints. It can’t violate those constraints or act freely outside the Runtime environment. Agents invoke tools through the Universal Tool Schema (UTS) and the ADL Capability Contract (ACC), Runtime features that govern tool use and authorization. We’ll explore UTS in a forthcoming article.
Agents can communicate with external actors and with each other. Each Runtime instance has a dedicated Shepherd agent responsible for managing that instance and the agent lifecycle.
Our First Product
Our first product built on ADL and the ADL Agent Runtime is CodeFriend. It uses an ensemble of agents to provide software application management services for developers and enterprises. CodeFriend is not a coding agent and doesn’t directly write application code. It’s intended to manage software at the application and repository levels.
CodeFriend — Because Your Code Needs a Friend
CodeFriend is intended to take on the tedious parts of the software development lifecycle, freeing developers’ time for more important tasks. Its purpose is to enhance developers’ abilities, not replace them.
In the upcoming Beta 1, CodeFriend provides five services:
• Documentation — Analyzes and documents your software based on its existing code.
• Diagrams — Generates diagrams of your software’s architecture and operations.
• Reviews — Performs nine kinds of internal reviews on your codebase, including Architecture and Security reviews, and documents all findings based on facts.
• Tests — Can determine your application’s test coverage and generate sufficient tests to reach a level you specify.
• Journaling — Records all changes to the codebase over time.
CodeFriend currently supports Rust, Python, Java, and JavaScript application repositories, with more languages coming soon.
We’ll open our private beta test in the next few weeks and hope you’ll join us as we continue to develop CodeFriend.
Future Directions
CodeFriend Beta 1 focuses on the five features above. In the future, we’ll support many additional languages and features, including fitness functions, automatic refactoring, and code transpilation.
Coming Soon
Agent Logic is just getting started. We have many more features and ideas to make AI agents more capable and manageable. The important point is that these are not opposing goals: the runtime amplifies what a model can reliably accomplish while governing what it may do.
Our vision is a world where AI and humans work together, complementing and amplifying each other’s abilities. Our soft launch will take place over the next few weeks, with more articles on the topics introduced here. Watch for future announcements and the CodeFriend Beta!
Contact: daniel@agent-logic.ai
