Long-Lived "Service Agents" for Context Offloading

[QuinRiva]

The Problem

Current Orchestrator --> Subtask flows are ephemeral and linear. When a coding agent needs nuanced information (e.g., complex database stats, large codebase architecture), we currently have to dump that data into the main context. This causes Context Pollution, degrading the coding agent's attention.

The Proposal Introduce Persistent Service Agents. Unlike sub-tasks which "do a job and die," Service Agents can exist independently of the main session to serve as specialised Knowledge Bases.

Key Characteristics

1. Persistence: They maintain their own context history
2. Protocol: The Primary Agent (Coding) can send structured queries to the Service Agent and receive summarized reasoning, not just raw data.
3. Model Specialisation: Users can assign high-context/low-cost models (e.g., Gemini 3 Flash) to Service Agents to digest massive amounts of documentation or DB schemas.

Note on the example below - I mention MCPs, this proposal is not about MCPs (yes I know that there are other solutions to the inherent context bloat of MCPs), this proposal is about having an background agent provide necessary and nuanced context to the coding agent.

---

Example Scenario: The Database SME

1. Service Agent (Gemini Flash): Connected to a DB via MCP. The DB holds JSON records in it's rows, so there's some complexity to it's structure. It reads the schema and analyses data distribution/content. It holds 500k tokens of DB stats in context.
2. Primary Agent (Claude 4.5 Opus): Needs to write a performant query.
3. Flow: Primary Agent asks Service Agent: "I need to query table X. Given the data distribution, should I index column Y?"
4. Result: Service Agent uses its context to answer "No, column Y has low cardinality," saving the Primary Agent from processing that raw data itself.

---

This proposal obviously relies on some concept of agent management, which I believe is already in the works.

If the Roo team is keen to implement this, I'd be happy to help out - I have developed several multi-agent interaction patterns like this, but I'm a Python developer, not a TypeScript developer.