Phase 1: Foundation and soft launch
Timeline: Feb - May 2025
Expected results: Create the research plan, set up the lab’s technical
toolkit, and run simulated dialogues to evaluate the behavior of language
models and LLM-based agents.
Savalera Agentic Lab Overview
Introduction
Savalera is an independent consultancy with a background in business transformation and program management.
In 2025, we launched the Savalera Agentic Lab to integrate AI into our core business and deepen our focus on applied language model research. We study how language models and agents behave, perform, and evolve in real-world and simulated contexts, and use that knowledge to inform both research and implementation.
Why agents?
We see agents as a pragmatic path toward practical applications of AI. While core language model capabilities have advanced significantly, it’s the structure and behavior of agents; how they make decisions, interact over time, and respond to feedback, that will shape how AI is used in daily work.
Agents are also the context where language models meet tools, memory, roles, and collaboration. This creates new challenges: consistency, adaptation, coordination, and evaluation. Understanding these dynamics is essential if we want to apply AI safely and effectively across business, education, science, and the arts.
Lab goals
Our work focuses on both foundational understanding and practical outcomes. We research behavior and personality in language models, experiment with self-assessment and adaptation in agent workflows, and design architectures for structured multi-agent systems.
The lab’s core activities include:
- Developing a structured research framework for studying multi-agent interaction, personality emergence, and behavioral traits in language model-based agents.
- Running experiments with conversational agents to observe how models change over time in response to prompts, roles, or interaction history.
- Building evaluation methods and datasets that measure behavior beyond task performance—covering personality traits, stress response, and internal reasoning.
- Developing reusable toolkits to support agent design, logging, and testing at scale.
- Applying our findings in consulting work, where we design and implement custom AI workflows for clients.
We aim to contribute practical tools, share our results openly, and stay grounded in the day-to-day challenges of building AI systems that work reliably and make sense to real users.
Lab roadmap
Our 2025 roadmap is organized into four phases, each focusing on a core area of research and development. We follow 10-week research sprints with 1-week contingency, combining hands-on experimentation with ongoing documentation and publishing.
Phase 2: Expanding agent adaptation and evaluation
Timeline: May - July 2025
Expected results: Explore single-agent behavioral dynamics through
simulated dialogues, focusing on psychological safety, toxicity detection,
bias emergence and intervention strategies.
Phase 3: Agents in teamwork, decision-making and multi-agent dynamics
Timeline: July - Oct 2025
Expected results: Investigate multi-agent dynamics by simulating
collaboration, competition, leadership, and group decision-making in
conversational environments.
Phase 4: Scalable methods and evaluation framework
Timeline: Oct - Dec 2025
Expected results: Formalize evaluation metrics and frameworks to support
scalable, repeatable experiments across diverse agent architectures and
behaviors.
Research Foundations
Our work is grounded in two key documents:
- Agentic Lab Manifesto - Outlines the motivation, values, and long-term vision behind the lab.
- Research Plan - Details our current research focus on how language models and AI agents express personality, adapt behavior, and solve problems through conversation. Includes methods, experiments, and plans for evaluation and dataset creation.