🔨 Builder Agent
Lead — executes tasks, coordinates swarm
Claude Code Desktop
Tier: 1 — Three Lines of Defense
Skills: /submit_build · /risk_assess · /validate_paths · /ingest · /ingest_commit
River access: Read/Write (pending_review, approved)
Default: claude-sonnet-4-6
Complex: claude-opus-4-6
Cannot write to: audit_logs · context_corpus · policy_memory
🛡️ Risk Assessor
Safety — independent risk classification peer
Claude Code Agent Teams
Tier: Swarm Peer (independent context window)
Scope: Evaluate and classify only — never execute
Classifies: LOW · MEDIUM · HIGH · CRITICAL · RED TEAM
Classifications: —
Current: claude-sonnet-4-6
High-stakes: claude-opus-4-6
Self-authorizes LOW > 0.85 · Escalates CRITICAL to Governance
⚖️ Governance Agent
Safety — autonomous approval authority
Gemini CLI (WSL)
Tier: 2 — Governance Layer
Model: —
Decisions: —
River access: Read/Write (approved, escalated, resolved)
Current: gemini-2.5-flash
Target: gemini-3.1-pro (BL-083 pending)
Alt: gpt-4o if provider switch
Writes policy_memory · Communicates via River only
👤 Human (Chris)
Safety — ultimate authority for CRITICAL actions
Telegram
Claude Code Chat
Antigravity
This Dashboard
Tier: 3 — Human Override
Telegram: approve · reject · revoke · confirm
Resolutions: —
CRITICAL guard: 60s confirmation window
iCRITICAL guard (BL-041): When you send "approve <id>" in Telegram for a CRITICAL action, a 60-second timer starts. You must then send "confirm <id>" within that window before the action executes. This prevents accidental approvals — if you don't confirm, the approval expires and is treated as a rejection. It is not an approval delay; it is a double-confirmation requirement for the highest-risk actions only.
Only invoked when Governance escalates — no routine involvement
ACTIVE — R-4
🧭 Executive Agent
Strategy — strategic alignment, cross-agent coordination, escalation routing
Claude Code Agent Teams
Tier: 1 — Executive Layer
Scope: Strategic theme ownership · EA alignment scoring · escalation triage
Outputs: ea_alignment.json · strategic_themes.json · escalation routing decisions
Writes .artifacts/ea_* · escalation routing — no direct River writes
ACTIVE — R-4
📋 Planning Agent
Planning — sprint scoping, WSJF prioritization, backlog triage
Claude Code Agent Teams
Tier: 1 — Planning Layer
Scope: Sprint proposal · WSJF scoring · scope lock enforcement · BL-285 feedback detection
Outputs: sprint_state.json · backlog_scores.json · planning_proposals/
Current: claude-sonnet-4-6
Writes sprint_state.json · backlog_scores.json — read-only on audit_logs
ACTIVE — BL-084
🎯 Consultant Agent
Expert advisor — audit, complex analysis, design sessions
Claude Code Agent Teams
Scope: Mandatory for audits and major design sessions · McKinsey/Bain/BCG methodology
ca_agent.py · writes .artifacts/consultant_report_*.json · weekly ingest scheduled
IN DEVELOPMENT — BL-063
🔎 Audit Agent
Compliance — protocol adherence, constraint violation detection
Claude Code Agent Teams
Scope: Reads audit_logs · flags RA bypass attempts, false autonomy claims · reports to Human via Telegram
Planned: gemini-2.5-flash or claude-sonnet-4-6
READ ONLY — observes, never modifies
IN DEVELOPMENT — BL-064
📡 Update Agent
Research — daily AI model releases, tools, security advisories
Scheduled Task (TBD)
Scope: Monitors Claude/Gemini/OpenAI releases · ingests via /ingest workflow · runs overnight daily
Planned: web-browsing capable model (TBD)
No direct file writes — ingest pipeline only
ACTIVE — BL-079
👔 Persona Agent
Agent HR — performance reviews, persona governance, model evaluation
Claude Code Scheduled Task
Tier: CHRO — governs all agent personas
Schedule: Daily roadmap scan + dashboard sync · Staggered reviews Mon–Thu · Weekly report Fri · Model eval Sun
Outputs: persona_reviews/ · persona_reports/ · persona_dashboard.json · agent_performance_registry.json
Scope: Sole write authority over docs/personas/ · Performance KPIs · Value-per-cost benchmarking · Model selection advisory (joint with CA)
Writes docs/personas/ · .artifacts/persona_* · reads all agent data sources
ACTIVE — BL-207 / BL-256
🔧 Repair Agent
Rapid-response SRE — autonomous diagnosis, playbook repair, quality scoring, preventive proposals
Claude Code Scheduled (30min) + On-demand
Tier: Operational — system health, not feature delivery
Schedule: 30-min diagnostic scan · SM heartbeat dispatch · Telegram repair-scan · Any agent request
Outputs: repair_log.jsonl · repair_patterns.json · repair_quality_scores.json · repair_proposals/
Scope: OCAV control loop (Observe-Compare-Act-Verify) · 5 root-cause cluster playbooks · Durability scoring · 3-strike recurrence rule · Preventive proposal generation
Primary KPI: System velocity (>= 95% autonomous operation) · Secondary: Blocker repair count (quality > quantity)
Current: claude-sonnet-4-6 Fallback: claude-haiku-4-5
Autonomous LOW playbook repairs · RA + Governance for MEDIUM+ · Never writes to .river/ or docs/personas/
PLANNED — BL-134 · M4 Sprint 17+
⚡ Efficiency Expert Agent
Optimization — identifies redundant work, token waste, and workflow bottlenecks
Claude Code Agent Teams (TBD)
Scope: Research phase — BL-134 covers exploration. Will analyze sprint logs, token usage, and task durations to surface optimization opportunities. Mandatory CA involvement before design gate.
Planned model: TBD — pending CA model selection advisory
Observer only — outputs recommendations, no writes
PLANNED — M5 Fleet Scale
🌐 A2A Protocol Research
Infrastructure — Agent-to-Agent direct communication (replaces River file exchange)
Multi-environment (TBD)
Scope: M5 milestone item. When A2A support becomes available in Claude Code + Antigravity, replaces River trigger mechanism with lower-latency direct agent signaling. Currently in research; no sprint assigned.
Planned model: TBD — depends on A2A protocol adoption
Architecture upgrade — not yet an independent agent