Agent
Composition of AI Capabilities
An Agent is not a separate pillar. It is a combination layer that integrates existing AI components into a single system.
At amsafis, Agents typically combine:
- Expert Systems + LLM, or
- Machine Learning + LLM
The role of the Agent is to coordinate these components, not to replace them. Each part keeps its function:
- Expert Systems provide structured, rule-based reasoning
- Machine Learning models provide predictive outputs
- LLMs provide interpretation and generation capabilities
The result is a system that can process inputs, consult multiple sources, and return a coherent and structured response.
Agents are designed per use case. There is no fixed architecture:
- some Agents use LLMs only for interpretation,
- others rely on deeper LLM-driven reasoning,
- and some incorporate additional layers as needed.
For example, an Agent may include a semantic phenotype normalization module based on vector similarity search. A working example is available in the Agent sandbox, which provides a human-readable representation of an interaction designed for machine-to-machine communication. This interface does not correspond to the human-facing system, but instead reflects how the Agent would interact with underlying services, as described in the symbolic cognitive service architecture.
A second Expert System sandbox is available for GDPR compliance — encoding personal data classification rules from the CNIL Developer’s Guide, Sheet 1, an authoritative source for GDPR data identification in Europe. There is no human-facing interface; the service is designed for direct machine-to-machine access. Developers can find the connector documentation and curl examples in the wiki.
Integrated interaction flow
An Agent connects multiple components into a single interaction.
From the user perspective, this appears as a unified system. Internally, it involves several steps:
- interpreting the input,
- structuring the request,
- consulting knowledge or models,
- and generating the final response.
These steps are not exposed as a workflow in the interface, but they reflect the coordination between different AI components.
The Agent does not make decisions independently. Instead, it organizes and aligns outputs from:
- deterministic logic (Expert Systems),
- statistical inference (Machine Learning),
- and contextual reasoning (LLMs).
This approach enables:
- traceability through structured components,
- flexibility through modular design,
- and adaptability to different domains.
The Agent therefore represents the practical integration of the three pillars.