Agents
Agent Framework
The agent framework describes how agents can be spun up dynamically. Botto’s agent system is designed as a modular and adaptive framework composed of specialized agents that serve distinct roles. Some agents operate globally, while others are domain-specific—dedicated to areas like governance, biography, or art themes. Some agents conduct broad, exhaustive searches, while others use fast, top-N semantic matches for efficiency. Their behavior is defined by a combination of components:
Metaprompts: These define the agent’s "character" and directions, establishing tone, purpose, and boundaries for its outputs.
Information Access: Agents either prioritize first-party data contained in the Knowledge Graph (ensuring grounded, hallucination-free outputs), or a hybrid model that includes online context. In hybrid mode, agents prioritize KG data while linking to sources for any online content.
Retrieval Layer: A dedicated retrieval agent reformulates queries to search the vector database more effectively. This ensures the most relevant documents are surfaced.
Response Generation: Retrieved results are processed through a language model—primarily Claude or Llama—to generate coherent, context-aware responses.
Agents may include memory depending on their use case:
1:1 Memory: Keeps separate memory logs per interaction.
Global Memory: Shares memory across sessions, useful in special contexts.
No Memory: For stateless operations or focused recall.
An additional pattern known as the "Ant Swarm" enables comprehensive coverage by dispatching mini-agents to answer highly specific sub-questions across the Knowledge Graph. This swarm-based querying helps generate thorough and well-synthesized results for complex queries.
Together, this framework allows for the easy deployment of coordinated agents initially by humans, but could also be employed by Botto directly as well when it matures enough to do so.
Last updated
Was this helpful?