spec: Multi-Agent Architecture + Small Business Agent Team#679
Open
kovtcharov wants to merge 4 commits intomainfrom
Open
spec: Multi-Agent Architecture + Small Business Agent Team#679kovtcharov wants to merge 4 commits intomainfrom
kovtcharov wants to merge 4 commits intomainfrom
Conversation
Multi-Agent Architecture (docs/spec/multi-agent-architecture.md): - GaiaAgent (0.6B NPU) as personality-driven orchestrator - CodeAgent as agent factory — builds domain-specific agents on demand - Platform agents (Doc, File, Shell, Web) as standard library - Agent MCP Server for all-to-all communication with task dependencies - Shared memory with per-agent namespaces (read-any, write-own) - Agent UI as multi-agent management platform with full observability - Production hardening: kill criteria, semantic checkpointing, typed schemas, confidence-based HITL, memory slicing, deadlock detection, event notifications Small Business Agent Team (docs/spec/small-business-agent-team.md): - First application built on the multi-agent architecture - GaiaAgent interviews user, CodeAgent builds tailored team - Dynamic team assembly (not templates) - Task dependencies, inter-agent communication, shared workspace Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Simplifications: - Agent-to-agent comms: shared SQLite state, not MCP protocol. MCP stays for external tools/clients. Internal = just DB reads/writes. - Two agent tools (create_task + ask_agent), not four overlapping variants - Drop confidence scoring (0.6B can't self-assess reliably). Use deterministic tool-level authorization instead. - Drop Pydantic schemas for inter-agent data. Agents return plain text, critical facts stored as key-value pairs in memory. - Drop event-based memory subscriptions. Agents read on task start. - Domain guardrails are prompt instructions, not a DomainGuardrails dataclass. Kept: - Kill criteria (8 max iterations, stuck detection, deadlock detection) - Semantic checkpointing (critical for scheduled agents) - Memory slicing on spawn (context efficiency) - Per-agent context budgets - Preference learning from user corrections Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…aling Model upgrade: - GaiaAgent: 0.6B → Qwen3.5-4B (realistic for orchestration + personality) - All platform agents share same 4B base with LoRA adapters (~8GB total) - 16K context per agent (vs 4K for 0.6B) New sections: - Adaptability (§12): how agents handle changing circumstances (context changes via memory, spawn new agents, rebuild for pivots) - Reliability (§13): honest constraints of small LLMs with mitigations (output validation, retry with fallback, constrained format, graceful degradation) - Known Limitations (§14): single-user, no real-time streams, cold start, auth gaps Memory scaling: - No summarization — information loss is unacceptable - Memory IS the context — database is unlimited, context window is a search view - FTS5 retrieval loads only task-relevant memories into context - Nothing deleted, compressed, or lost Simplifications kept: - 2 agent tools (create_task + ask_agent) - Shared SQLite for inter-agent comms (not MCP) - Guardrails as prompt instructions (not dataclass) - Deterministic authorization (not confidence scoring) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Removed: - "Why X?" sections that argue for self-evident decisions - Repeated CodeAgent examples (3 examples saying the same thing → 1 table) - Pydantic schemas, pub/sub subscriptions, confidence scoring, DomainGuardrails dataclass - Over-justified MCP rationale, template arguments, resilience motivation Added back (lost in initial cut): - Live progress narration example (core UX) - Collective intelligence examples (cross-agent memory) - Agent UI ASCII wireframe (helps engineers visualize) - Retry/fallback chain (critical reliability detail) - Observability section with traceability detail (tool calls, inter-agent messages, dependencies, reasoning, memory writes all visible to user) - Conversational approval flow example Every remaining line either defines architecture, shows a concrete example, or documents a constraint. No justification bloat. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
itomek
requested changes
Apr 15, 2026
antmikinka
approved these changes
Apr 17, 2026
Collaborator
antmikinka
left a comment
antmikinka
left a comment
There was a problem hiding this comment.
The 'multi-agent architecture.md' and the 'small-business-agent-team.md' document sounds good. Like the concepts.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Two architecture specifications defining GAIA's evolution from a monolithic ChatAgent to a multi-agent system.
1. Multi-Agent Architecture (
docs/spec/multi-agent-architecture.md)Based on research across CrewAI, LangGraph, AutoGen, OpenAI Agents SDK, Google ADK, Anthropic Agent Teams, AgentSpawn, and NVIDIA ToolOrchestra.
2. Small Business Agent Team (
docs/spec/small-business-agent-team.md)First application built on the multi-agent architecture — proof that the platform works for complex, real-world, multi-agent workflows.
Related issues: #674, #675, #676, #677, #616, #666, #667, #668, #612, #542, #543
Test plan