|
Great questions — and yes, you’ve got the right intuition: we orchestrate role-specific agents using off-the-shelf LLMs (Claude 3.7, DeepSeek GPT 4.1, GPT-4 Turbo), and they “hand off” tasks between each other. But to avoid total chaos (or hallucinated collaboration), we had to build a few things around them. The “shared context layer” is essentially a lightweight memory and coordination layer that persists project state, intermediate decisions, and validated outputs. It’s not a traditional vector store or RAG setup. Instead, we use:
• A Redis-backed scratchpad with typed slots for inputs, constraints, decisions, outputs, and feedback
• An MCP (Model Context Protocol) template that defines what agents should expect, expose, and inherit
• Each agent works statelessly, but gets a structured payload that includes relevant validated history, filtered to reduce noise Agents don’t have full access to each other’s output logs (too much context = hallucination risk). Instead, each one produces an “artifact” + optional feedback object. These go into the shared layer, and the orchestrator decides what the next agent should receive and in what form. We don’t run anything locally (yet). It’s all API-based for now, with orchestration handled in a containerized layer. That will probably evolve if we scale into more sensitive verticals. Hope that helps clarify. Happy to dig deeper if you want building something similar. |