Multi-Agent Today's Tasks System - Implementation Plan

Date: 2025-03-01 Status: Architecture Plan Target System: Today's Tasks Workflow (Multi-Agent Committee)

📋 Executive Summary

This document outlines the implementation plan for transforming the "Today's Tasks" system from a single-prompt generator into a Multi-Agent "Committee" Architecture.

Instead of a generic LLM generating tasks, we will leverage our existing specialized agents (StrategyArchitect, ContentStrategist, SEOOptimization, etc.) to propose high-value, context-aware tasks based on their specific domain knowledge. A central "Manager" (Orchestrator) will then consolidate, prioritize, and deduplicate these proposals into a cohesive daily plan.

We will also introduce a Self-Learning Task Memory using txtai to ensure the system learns from user behavior (acceptances/rejections) and avoids redundant suggestions.

🏗️ Architecture: The "Committee" Model

1. Agent Roles & Responsibilities

Each agent will act as a "Department Head," submitting daily proposals for their specific pillar.

Workflow Pillar	Owner Agent	Data Sources	Proposal Type Example
PLAN	`StrategyArchitectAgent`	Content Pillars, Strategy Doc	"Review 'AI Trends' pillar strategy - engagement dropped 10%."
GENERATE	`ContentStrategyAgent`	Content Gaps, Trends	"Draft a blog post on 'Vector Search' (High Opportunity Gap)."
PUBLISH	`SocialAmplificationAgent`	Audience Activity, Calendar	"Schedule your 'Weekly Recap' thread for 10 AM (Peak Audience)."
ANALYZE	`SEOOptimizationAgent`	GSC, Site Health, Rankings	"Fix 3 broken links on your pricing page to recover link equity."
ENGAGE	`SocialAmplificationAgent`	Social Mentions, Comments	"Reply to 3 unanswered comments on your latest LinkedIn post."
REMARKET	`CompetitorResponseAgent`	Competitor Activity	"Competitor X posted about [Topic]. Create a counter-narrative Reel."

2. The Workflow (Daily Cycle)

Morning Briefing (Parallel): TodayWorkflowManager polls all agents via propose_daily_tasks(context).
Aggregation: Manager collects raw proposals (~10-15 tasks).
Intelligence Filter (Self-Learning):
- Check TaskMemoryIndex (txtai).
- Filter out tasks similar to previously Rejected tasks.
- Deprioritize tasks similar to recently Completed tasks.
Consolidation: Deduplicate overlapping ideas (e.g., SEO & Content agents both suggesting the same topic).
Final Selection: Select top 1-3 tasks per pillar based on user goals (e.g., "Growth" mode = more Publish/Remarket tasks).

🚀 Implementation Phases

Phase 1: Agent Interface Standardization (The "Voice")

Objective: Give every agent the ability to speak the "Task Proposal" language. Status: ✅ Completed

Task 1.1: Define TaskProposal schema (Pydantic model). ✅
- Fields: title, description, pillar, priority, reasoning, estimated_time, action_type.
Task 1.2: Update BaseALwrityAgent with abstract propose_daily_tasks(context: Dict) -> List[TaskProposal]. ✅
Task 1.3: Implement propose_daily_tasks in all specialized agents. ✅
- StrategyArchitect: Logic to check pillar health.
- ContentStrategist: Logic to check content gaps.
- SEOAgent: Logic to check GSC alerts/errors.
- SocialAmplification: Logic for publish/engage.
- CompetitorResponse: Logic for monitoring.

Phase 2: The Manager (The "Orchestrator")

Objective: Build the backend service that coordinates the committee. Status: ✅ Completed

Task 2.1: Refactor TodayWorkflowGenerator in today_workflow_service.py. ✅
- Replace single-prompt generation with gather_agent_proposals().
- Implement asyncio.gather for parallel agent execution (performance critical).
Task 2.2: Implement consolidate_proposals() logic. ✅
- Use a lightweight LLM call to merge/rank the raw list if needed, or deterministic logic for speed.
Task 2.3: Connect to Frontend. ✅
- Ensure the API response matches the existing TodayTask frontend interface.

Phase 3: Self-Learning Memory (The "Brain")

Objective: Stop the system from nagging users about things they hate or just did. Status: ✅ Completed

Task 3.1: Create TaskHistory model in DB. ✅
- Store: task_vector_id, original_text, status (completed/rejected/skipped), user_feedback.
Task 3.2: Implement TaskMemoryService using txtai. ✅
- Index tasks with metadata.
- Implement is_redundant_or_rejected(proposal_text) check.
Task 3.3: Wire feedback loop. ✅
- When user clicks "Dismiss" or "Complete" in frontend, update the txtai index.

Phase 4: UI Feedback & Transparency

Objective: Show the user why a task was suggested.

Task 4.1: Update Frontend TodayTask card.
- Add "Suggested by [Agent Name]" badge.
- Add "Why?" tooltip (e.g., "Because Competitor X did Y").
Task 4.2: Add "Train my Agents" feedback.
- "Don't show this again" vs "Not today".

📊 Data Models

1. TaskProposal (Backend)

class TaskProposal(BaseModel):
    title: str
    description: str
    pillar_id: str  # plan, generate, publish, analyze, engage, remarket
    priority: str   # high, medium, low
    estimated_time: int # minutes
    source_agent: str # e.g., "SEOOptimizationAgent"
    reasoning: str # "Detected 404 error spike"
    context_data: Optional[Dict] # e.g., {"url": "..."}

2. TaskMemoryDocument (txtai)

{
  "id": "uuid",
  "text": "Write a blog post about AI Trends",
  "embedding": [vector],
  "tags": ["generate", "content_strategy"],
  "user_id": "123",
  "status": "rejected",
  "last_updated": "2024-03-01T10:00:00Z"
}

🛠️ Technical Considerations

Performance: Calling 6 agents + LLMs in parallel can be slow.
- Mitigation: Set strict timeouts (e.g., 5s) per agent. Use "Lite" logic for proposals (e.g., check DB/Cache instead of live crawling) where possible.
Fallback: If agents time out or fail, fall back to the _fallback_tasks template currently in place.
Token Usage: Summarize context before sending to agents to minimize input tokens.

📅 Execution Timeline

Day 1: Phase 1 (Interfaces & 2 core agents).
Day 2: Phase 2 (Orchestrator wiring).
Day 3: Phase 3 (txtai Memory integration).
Day 4: Phase 1 completion (remaining agents) & Phase 4 (UI polish).

6.6 KiB Raw Blame History