I've kept quite busy over the last few months, building what I hope will become a valuable tool and platform. While we're still a few weeks away from a potential big reveal , the signs are promising, and progress has been swift. At the core of this solution is an agentic framework-an overarching service layer designed as a dedicated, highly optimised, and extensible expert system. Its primary aim is to transform data into actionable insights and provide knowledge and actions plans that improve outcomes. Agentic frameworks have been a significant focus of mine for some time.
In recent months, they've started to emerge not only as abstract patterns but as tangible solutions. As these frameworks develop, I'm noticing some interesting design choices. While these designs often hit the target, they sometimes miss the bullseye of what agentic frameworks can truly deliver. Here are three key principles I believe form the foundation of a strong agentic framework: 1. Context Frameworks must have the ability to share and enrich context. They should learn from their actions and act not only as data receivers, but as data transmitters, discovering and integrating new data sources seamlessly.
Context isn't static; it evolves and should inform decision-making at every level of the framework. 2. Authority Agents within these frameworks must provide expertise and domain-specific knowledge. Relying solely on generic large language model (LLM) agents performing sequential transactional tasks is insufficient. Instead, each agent should be purpose-built, with predefined and finely tuned capabilities. This ensures optimised routing and authentic responses, enabling the system to act with confidence and precision.
3. Outcomes Outcomes should drive the framework-both as inputs and outputs. The steps required to achieve these outcomes should be dynamically generated, always focusing on the most efficient and effective paths. Orchestration is critical: there must be a controller/orchestration agent whose role is to continually validate whether outcomes have been achieved and to optimise the workflow dynamically for better results.