Multi-Agent Architecture for GAIA Benchmark

Goal

Build a multi-agent architecture that can reliably solve GAIA-style benchmark questions through agent collaboration, shared workspace, and orchestrated handoffs.

Architecture Overview

Implement a 3-agent system inspired by Manus: 1. Planner Agent - analyzes the question, breaks it into steps, creates execution plan 2. Executor Agent - executes the plan using available tools (web search, code execution, etc.) 3. Verifier Agent - reviews results, checks answer quality, validates against question requirements

Core Components

Shared Workspace

• Task state (current step, progress, artifacts)
• Agent communication log (handoffs, questions, findings)
• Intermediate results storage (search results, extracted data, etc.)
• Final answer candidate

Agent Communication Protocol

• Explicit handoff messages between agents
• State transitions (planning → executing → verifying → done)
• Ability to loop back if verification fails
• Guardrails against infinite loops (max iterations)

Orchestration Logic

• Spawns agents in sequence: Planner → Executor → Verifier
• Manages state transitions
• Handles agent failures/retries
• Aggregates final result

Test Case

GAIA Level 1 Question: "What was the actual enrollment count of the clinical trial on H. pylori in acne vulgaris patients from Jan-May 2018?"

Expected behavior: 1. Planner breaks down: need to search clinical trials database → filter by criteria → extract enrollment count 2. Executor performs web searches, navigates to trial details, extracts data 3. Verifier checks: is this the right trial? does it match date range? is enrollment count clearly stated? 4. Return exact numeric answer

Success Criteria

• Architecture can solve the test case question correctly
• Agent handoffs are clean and traceable
• Shared workspace maintains clear state
• System fails gracefully (doesn't infinite loop or hallucinate)

Implementation Notes

• Use existing subagent spawning infrastructure
• Shared state can be a simple dict/object passed between agents
• Each agent gets a specialized system prompt defining its role
• Consider adding a "max_rounds" guardrail (e.g., 3 full cycles max)

Namespace

Omni/Agent/MultiAgent