Committee System

The agent committee is the core of ALwrity's daily workflow generation. Six specialised agents are polled in parallel to propose tasks across the 6 Content Lifecycle Pillars.

The 6 Pillars

Pillar	Focus	Accepts Tasks From
plan	Content strategy & planning	Content Strategy, Strategy Architect
generate	Content creation	Content Gap Radar
publish	Content distribution	All agents
analyze	Performance analysis	SEO Optimization, Competitor Response
engage	Social engagement	Social Amplification
remarket	Content repurposing	All agents

Polling Flow

Today's Workflow Generation
        │
        ▼
1. Build Grounding Context
   └── Onboarding data + unread agent alerts
        │
        ▼
2. Poll Committee (parallel)
   ├── Content Strategy Agent
   ├── Strategy Architect Agent
   ├── SEO Optimization Agent
   ├── Social Amplification Agent
   ├── Competitor Response Agent
   └── Content Gap Radar Agent
   └── Each: propose_daily_tasks(context) → List[TaskProposal]
        │
        ▼
3. Deduplication
   └── Remove exact title+pillar duplicates (priority-based tiebreaking)
        │
        ▼
4. Self-Learning Filter
   └── TaskMemoryService.filter_redundant_proposals()
       ├── Remove exact hash matches from last 7 days
       └── Remove semantically similar (txtai > 0.85) to dismissed tasks
        │
        ▼
5. Pillar Coverage Enforcement
   └── Backfill missing pillars via LLM-generated tasks
   └── Controlled fallback if LLM fails (template tasks)
        │
        ▼
6. Committee Watchdog Audit
   └── ContentGuardianAgent.audit_committee(proposals)
       ├── Per-agent critique (reasoning, priority, pillar fit, acceptance rate)
       ├── Coverage gap detection
       ├── Overlap detection
       └── Alert generation for serious faults
        │
        ▼
7. LLM Fallback (if committee returns nothing)
   └── Generate all 6 pillars via llm_text_gen()
        │
        ▼
8. Contextuality Validation
   └── Each task must have ≥1 evidence link to onboarding or alerts
   └── Score threshold: 0.65
        │
        ▼
9. SIF Indexing (fire-and-forget)
   └── Tasks indexed into txtai for semantic search

Deduplication

When two agents propose the same or similar tasks, the system resolves by priority:

• Same title + same pillar → keep the one with higher priority (high > medium > low)
• Same title + different pillar → both kept (different execution contexts)
• Semantic duplicates (txtai similarity > 0.85) within 7 days of a dismissed task → removed

Pillar Coverage Enforcement

After deduplication, any pillar with zero tasks triggers LLM-based backfill:

for pid in PILLAR_IDS:
    if pid not in covered_pillars:
        llm_task = generate_task_for_pillar(pid, context)
        if llm_task:
            tasks.append(llm_task)

If the LLM call fails, hardcoded template tasks are used as a fallback.

Contextuality Validation

Each task is scored against the grounding context (onboarding data + agent alerts). A task must have at least one evidence link — a reference to specific user data or an unread alert — to pass. If the plan's average score is below 0.65, the system regenerates with strict contextuality enforcement.

Committee Engine Code

The committee logic lives in backend/services/today_workflow_service.py:

Function	Lines	Responsibility
generate_agent_enhanced_plan()	391–607	Main engine — polls agents, deduplicates, validates, runs audit
build_grounding_context()	341–381	Aggregates onboarding data + unread alerts
_ensure_pillar_coverage()	305–338	Backfills missing pillars
validate_plan_contextuality()	197–254	Scores plan quality against evidence
_fallback_tasks()	50–112	Hardcoded fallback tasks

Agent-to-Pillar Mapping

The committee validates proposals against expected pillar assignments:

Agent	Expected Pillar Focus
ContentStrategyAgent	plan
StrategyArchitectAgent	plan
SEOOptimizationAgent	analyze
SocialAmplificationAgent	engage
CompetitorResponseAgent	analyze
ContentGapRadarAgent	generate

When an agent proposes outside its expected pillar, the ContentGuardianAgent flags it as an "off-pillar" issue.