CabraMesh · Agentic engineering
CabraMesh is the name of our V2 research in agentic engineering: a local, multi-agent agentic architecture, still in design and separate from the open-source rf-test-agents solution (V1). Useful AI must do more than propose. It must act within controlled boundaries, under human oversight, and support its conclusions with evidence.
More autonomy, without losing control
AI Cabra is designing a local agentic architecture for demanding enterprise environments. Our direction: combine model flexibility with software engineering discipline, without mistaking a convincing answer for a verified result.
People remain in control
Objectives and boundaries remain under your control. The intended autonomy stays within an approved scope, not an unlimited delegation.
Your data stays with you
The design prioritizes controlled local infrastructure, with no implicit dependency on an external service to process your private data.
Verifiable results
Quality is assessed through observable facts. Actions and their outcomes must be open to review, discussion and validation.
From an assistant that answers to a team that moves work forward
Imagine agents that can prepare work, test it against reality and have the outcome independently reviewed, without taking the decision away from you. Our research aims for that continuity: move from business intent to controlled actions, with clear responsibilities and inspectable evidence. The STDD method provides the starting point: reality before promises.
Turn intent into concrete work
Start from a business need, clarify expectations and prepare an approach your team can use. The direction: fewer round trips between an idea, a plan and putting it to the test.
Evolve your AI without losing your boundaries
Choose models to suit the work while retaining your rules, limits and quality expectations. The model should serve your organization, not impose its way of working.
Automate without turning a blind eye
Aim for traceable actions and verified outcomes, rather than a success message alone. When something diverges, retain the ability to understand, decide and correct within an approved scope.
Four engineering choices behind autonomy
STDD is a method that uses agents to test specifications against reality; it is not an agentic architecture. rf-test-agents is the open-source tooling that applies this method. V2 opens a separate line of work: engineering an architecture that separates reasoning, authorization, execution and verification. These are the technical directions of our research, without exposing internal protocols.
Specialized roles, bounded contexts
Planning, acting and verifying require different information and permissions. V2 aims for bounded missions, role-appropriate context and a review separate from producing the outcome. An agent should not decide on its own what it is allowed to do.
Models qualified for each task
A model suited to preparing a plan is not necessarily suited to analysing a discrepancy. We are exploring a local multi-model architecture: select capabilities using representative cases, then compare quality, latency and resource requirements. Changing a model should not change business rules.
Execution governed by deterministic checks
The model proposes an action; software checks must verify its scope before execution. V2 aims for a clear distinction between read access, changes and human approval, with explicit tool interfaces. A natural-language instruction is not, by itself, authorization.
Evidence beyond the agent’s verdict
A “done” message is not enough. The aim is to connect the expected outcome to system observations, changes made and checks performed. Acceptance criteria remain the reference; failure must stay visible, without weakening assertions to obtain a green result.
V1 · The open-source rf-test-agents foundation
Version: 1.0.0 · Apache 2.0
The original rf-test-agents solution is available as open source under the Apache 2.0 license, for personal, commercial or industrial use. It brings together five specialized agents to plan, generate, heal, verify and document Robot Framework tests, using the STDD method and its safeguards.
V1 can be used now. Our V2 work builds on this foundation and explores a new way to bring together models, actions and verification. It remains under research and is not part of the available open-source version.
CabraMesh · Evolving our technologies and our thinking
Our work on test agents leads us to broaden the question: how can different capabilities work together without losing business purpose, human control or the quality of outcomes? V2 explores that next step. It evolves our technologies as well as the way we think about autonomy: models supporting a method, actions within clear boundaries and verification retaining its full role.
What boundaries should govern your AI agents?
Start with your business context, data constraints and expected level of control to scope the next steps.
- A technical exchange, not a sales pitch
- Your data stays in France