Agent eXecution Model — governed execution, built around the problem to solve.
AXM develops an execution model for work carried out by AI agents, software and people. It starts with the problem, the result expected and the conditions for accepting it. The public starting point is AXM Forge, a Python developer toolchain available through MCP, a shared CLI and Python libraries.
Think of commissioning a piece of work in a workshop. The brief and acceptance criteria remain attached to the job as different specialists contribute. Changing who does the work need not change what counts as an acceptable result.
AXM applies that separation to execution. The problem and its progress persist across attempts. A contract describes what an acceptable result must satisfy; an evaluation policy describes how to assess it; an executor attempts the work. Evaluation can combine automated checks, model judgment and human review, according to the task.
The design principle is to make the decision to accept a proposed change explicit and inspectable. This lets the means of execution evolve while keeping the requirements visible. The strength of an acceptance decision depends on the checks and evidence behind it.
The blog explores the ideas behind this approach: constraints within autonomous loops, bounded execution and the difference between a convincing result and a justified one.
Forge provides tools for understanding and changing code:
- Code intelligence — structural analysis, symbol search and dependency exploration.
- Quality checks — lint, types, tests, coverage, security and project conventions.
- Editing and refactoring — batch edits with validation and rollback support, plus Python symbol refactoring.
- Developer workflows — scaffolding, Git operations and text compaction.
Tools return structured results with a compact text representation for agent clients. Install the providers you need; the available catalog depends on the server environment.
For MCP, start with the installation and first-call guide. It covers a client-managed stdio process and a persistent Streamable HTTP server. For individual tools and Python libraries, browse the package catalog.
Private development explores how to carry that model through complex work: compose units of work, recover from failed attempts and bring unresolved decisions back to a person. The aim is to reduce the attention needed to supervise execution while preserving a clear account of its progress.
Those components are not offered here as public packages. This profile introduces the direction; the public repositories and their documentation define what is available to use today.
Forge includes tooling to check both code quality and project conventions. Its CI and package documentation expose the relevant checks and usage limits.
Forge is licensed under Apache 2.0. Licensing and availability are specific to each repository; private development is not a commitment to a public release.
