Agent Registry System Hub
The single lookup table for all 44 agent directories under ~/.openclaw/agents/ — covering 42 live agent configs (2 strategic Opus, 14 operations Sonnet, 18 automation Haiku, 8 special/sub-agent/inactive) plus 2 infrastructure dirs (_retired/, _templates/). Read this hub before assigning any task to an agent, modifying an agent config, adding a new service, or checking domain ownership. The canonical doc AGENT-REGISTRY is stale (last updated 2026-04-07 — 26 days behind live state as of 2026-05-03). This hub supersedes it for navigation and cross-link purposes.
Quick reference
| Field | Value |
|---|---|
| Canonical file | /home/opsadmin/.openclaw/workspace/AGENT-REGISTRY.md |
| Live state probe | ls ~/.openclaw/agents/ | wc -l |
| Documented count | 44 dirs (CLAUDE.md, §“44 agents”) |
| Live count | 44 dirs (probed: 2026-05-03) |
| Drift | 0% (dir count matches) |
| Last audit | 2026-05-03 |
| Drift alert | manual (no cron yet — Phase 1.5 pending) |
| Governance gate | G-SERVICE-PRE-START-DOC (new agents must be registered before first start) |
| Owner agent | _summary |
⚠️ AGENT-REGISTRY.md staleness alert
/home/opsadmin/.openclaw/workspace/AGENT-REGISTRY.md was last updated 2026-04-07 — 26 days behind this audit (2026-05-03). Key known gaps vs current live state:
- AGENT-REGISTRY.md documents 45 agents (the 2026-04-07 reconciliation count); current CLAUDE.md canonical count is 44 dirs (42 active + 2 infra dirs).
- The registry still lists
betteracqanddealcmdas “RETIRED” inline but they appear in the live/home/opsadmin/.openclaw/agents/dir as active configs — confirmed:betteracqanddealcmddirs exist in_retired/subdirectory as of 2026-05-03 audit. stefania,stefania-asst,defaultwere retired 2026-04-24 per CLAUDE.md; their entries remain in the registry with “Sub-agent” status rather than “Retired”._retired/contains 16 subdirs (angel-calls, betteracq, codex, deal-command-center, dealcmd, debugger, default-retired-2026-04-24, h2-calls, h3-calls, investorlift-angel, investorlift-h2, investorlift-h3, stefania-asst, stefania-asst-retired-2026-04-24, stefania-retired-2026-04-24, trackcommander) — none listed in registry.
Action required: Update AGENT-REGISTRY.md to reflect current state. Until then, use this hub + CLAUDE.md as source of truth.
Live state snapshot (2026-05-03)
| Metric | Documented | Live (probed) | Drift | Status |
|---|---|---|---|---|
Total dirs in ~/.openclaw/agents/ | 44 | 44 | 0% | ok |
Non-system agent dirs (excl _retired, _templates) | 42 | 42 | 0% | ok |
_retired/ subdirs | ~3 (stefania, stefania-asst, default) | 16 | +433% | drifted — many more retired agents than documented |
_templates/ items | 1 | 1 | 0% | ok |
| Strategic tier (Opus) | 2 | 2 | 0% | ok |
| Operations tier (Sonnet) | 14 | 14 | 0% | ok |
| Automation tier (Haiku) | 18+ | 18 | 0% | ok |
| Special/sub-agent/inactive | 8 | 8 | 0% | ok |
| Vault wiki/agents summary pages | 42 | 42 | 0% | ok (all agents have _summary.md in vault) |
| AGENT-REGISTRY.md freshness | current | 26 days stale | — | flagged |
Agent catalog
Strategic Tier — Opus 4 (2 agents)
Heavy reasoning. Use for architecture, governance, novel problem-solving. Model: claude-opus-4-20250514.
| Agent | Summary page | Purpose | Domain | Escalates to |
|---|---|---|---|---|
| aurora | _summary | Chief AI Ops, Discord commander, primary builder | Platform ops, Discord, build orchestration | solara |
| solara | _summary | Strategic orchestration, system governance | Compliance, risk, SYSTEM-STATE.md, P0 plans | Henry |
Operations Tier — Sonnet 4 (14 agents)
Balanced, reliable. Use for API integrations, data pipelines, multi-step workflows. Model: claude-sonnet-4-20250514.
| Agent | Summary page | Purpose | Domain | Key service(s) |
|---|---|---|---|---|
| main | _summary | Central orchestration, 80%+ of cron jobs | Cross-agent coordination, SM token refresh | sm-token-refresh.service |
| atlas | _summary | Enterprise architecture, InvestorLift ops | InvestorLift, property data, Docker | property-view-watcher.service, investorlift-gateway.service |
| acquisitions | _summary | Acquisitions engine, inbound lead handling | HubSpot inbound leads, deal intake | — |
| betterfiles | _summary | Document management, TC coordination | Gmail labels, DocuSign | — |
| bizintel | _summary | Business intelligence, market analysis | BI dashboards, market data | — |
| dispo | _summary | Disposition management | Dispo pipeline, buyer outreach | — |
| reri | _summary | Real estate operations (RERI brand) | RERI brand workflows | — |
| ryan-norcal | _summary | Sales ops — Northern California | NorCal lead management | — |
| david-socal | _summary | Sales ops — Southern California | SoCal lead management | — |
| hhill-socal | _summary | Founder/executive operations | Exec-level workflows, strategy | — |
| auctionscout | _summary | Auction intelligence and monitoring | Auction data, bid analysis | — |
| prediction-trader | _summary | Prediction market + crypto trading | Kalshi, Polymarket, crypto arbitrage | prediction-ingestor.service, prediction-history.timer |
| betterbets | _summary | Multi-sport betting orchestrator | Horse racing, sports betting | — |
| bettertrading | _summary | Trading automation | Crypto/market automation (inactive) | — |
Automation Tier — Haiku 4.5 (18 agents)
Fast, cheap, high-volume. Use for polling, routine data sync, repetitive tasks. Model: claude-haiku-4-5-20251001.
| Agent | Summary page | Purpose | Domain |
|---|---|---|---|
| research | _summary | Web research, data gathering | External data fetching, scraping |
| ops | _summary | Operations automation | Cron-driven automation, backup |
| mktg | _summary | Marketing automation | Campaign management |
| webops | _summary | Web operations | Web scraping, monitoring |
| social | _summary | Social media management | Social posting, monitoring |
| backfill | _summary | Bulk data backfill | HubSpot/SalesMsg contact backfills |
| bettershowings | _summary | Better Showings coordination | Showing scheduling |
| betterhomeoffers | _summary | Home offer automation | Offer generation |
| bettercredit | _summary | Credit-related operations | Credit workflows |
| betterarbitrage | _summary | Arbitrage opportunity finder | Market arbitrage detection |
| tradelines | _summary | Tradelines management | Tradeline operations |
| crowdfunding | _summary | Crowdfunding operations | Crowdfunding workflows |
| fund-calls | _summary | Fund call coordination (parent) | Fund investor call orchestration |
| angel-calls | _summary | Angel investor call sub-variant | Angel investor outreach |
| h2-calls | _summary | H2 investor call sub-variant | H2 investor outreach |
| h3-calls | _summary | H3 investor call sub-variant | H3 investor outreach |
| investorlift | _summary | InvestorLift integration (parent) | IL property matching, data sync |
| investorlift-angel | _summary | InvestorLift angel entity sub-variant | Angel investor IL matching |
Note:
investorlift-h2andinvestorlift-h3are listed as Haiku sub-variants in AGENT-REGISTRY.md but are categorized here as automation tier for consistency. Total Haiku count is 18 when counting only distinctly-owned agents (investorlift-h2 and investorlift-h3 included but pushed to the sub-agent group below for clarity).
Sub-agents and special configs (8 configs)
| Agent | Summary page | Status | Notes |
|---|---|---|---|
| investorlift-h2 | _summary | Active | H2 entity IL matching sub-variant |
| investorlift-h3 | _summary | Active | H3 entity IL matching sub-variant |
| kimi | _summary | Special | Moonshot Kimi model for long-context tasks |
| chatgpt | _summary | Special | GPT-4o bridge for OpenAI-specific tasks |
| anthropic | _summary | Config-only | Anthropic direct provider config (not a real agent) |
| tablegames | _summary | Inactive | Table games analysis — paused |
| stefania | — | Retired 2026-04-24 | Haiku persona config. Handling transferred to Aurora per project_stephania_aurora_chat_wired |
| stefania-asst | — | Retired 2026-04-24 | Auth-profiles only. Sessions dir preserved in _retired/ |
| default | — | Retired 2026-04-24 | Catch-all session bucket. Sessions dir preserved in _retired/ |
Previously retired (merged earlier)
| Agent | Status | Notes |
|---|---|---|
| betteracq | Retired | Merged into acquisitions agent |
| dealcmd | Retired | Merged into dispo agent |
Components
/home/opsadmin/.openclaw/workspace/AGENT-REGISTRY.md— primary registry doc (⚠️ stale: 2026-04-07)/home/opsadmin/.openclaw/agents/<name>/agent/— per-agent config dir (models.json, SOUL.md, TOOLS.md)/home/opsadmin/.openclaw/agents/<name>/agent/SOUL.md— agent identity (read BEFORE modifying any agent — G-AGENT-IDENTITY-FIRST)/home/opsadmin/.openclaw/agents/<name>/agent/models.json— model routing config (⚠️ overwritten by gateway on restart — edit viaportkey-proxy/models-backups/per feedback_models_guardian_pattern)/home/opsadmin/.openclaw/agents/_retired/— 16 retired agent dirs preserved per feedback_archive_not_delete/home/opsadmin/.openclaw/openclaw.json— gateway config; master list of active agent routes/home/opsadmin/openclaw-vault/wiki/agents/— vault mirror; 42_summary.mdpages exist for all active agents
How it’s used
- Trigger: before assigning any task, modifying an agent config, adding a service, or checking ownership — consult this hub and AGENT-REGISTRY.md
- Workflow: request → AGENT-REGISTRY lookup → domain owner identified → SOUL.md read → task assigned → WORKSPACE-REGISTRY for path → proceed
- Agents involved: _summary (primary builder/assigner), _summary (governance oversight)
- Failure mode: assigning tasks to wrong-domain agents causes work in wrong workspace paths, conflicting script ownership, and service restarts that kill each other’s processes. Past incident: 2026-04-16 Aurora SOUL overwrite (assigned wrong-layer capability). See feedback_agent_identity_first.
- Success criteria: every task has exactly one owning agent identified before first tool call; service restarts are pre-authorized against the correct owner
Agent tiers and model routing
All LLM calls route through portkey proxy at 127.0.0.1:18900.
| Tier | Model ID | Agents | Max plan? |
|---|---|---|---|
| Strategic | claude-opus-4-20250514 | aurora, solara | yes (henryRERI) |
| Operations | claude-sonnet-4-20250514 | 14 Sonnet agents | yes (teamsteph@betterfiles.com via :18903) |
| Automation | claude-haiku-4-5-20251001 | 18+ Haiku agents | via anthropic-max-router |
Model ID expiry: old IDs deprecated June 15 2026. Monitor for breaking changes.
Cross-links
Agents that own this hub
- _summary — authors and maintains registry
- _summary — governance oversight (tier / retirement decisions)
Skills that invoke agent registry checks
- il-marketplace-pull — reads atlas domain assignment
- hubspot-deal-ingest — reads atlas domain assignment
- acquisitions-outreach — reads acquisitions domain assignment
- acquisitions-followup — reads acquisitions domain assignment
- dispo-blast — reads dispo domain assignment
Plans that govern agent lifecycle
- openclaw-fragmentation-fix-2026-05-01 — Phase D: agent identity drift resolution
- openclaw-self-improvement-layer-2026-05-03 — OSIL B10: LiveKit voice substrate agent tier
- nemoclaw-audit-2026-05-03 — proposed L7 credential injection pattern affecting all agents
Feedback rules
- feedback_agent_identity_first — read SOUL.md before any agent change
- feedback_models_guardian_pattern — models.json overwritten by gateway; edit backups instead
- feedback_action_gate_violation_repeated — all service restarts require explicit auth
- feedback_archive_not_delete — retired agents →
_retired/, never deleted
KB / source docs
- API — gateway API reference
- AGENT-REGISTRY — stale source (2026-04-07); superseded by this hub for navigation
System maps
- agents-tier-structure — visual tier diagram (99 incoming wikilinks — stealth hub candidate)
- request-lifecycle — how requests route through tiers (50 incoming)
Related hubs
- skill-registry — skills owned by each agent domain
- service-registry — systemd services owned by each agent
- workspace-registry — workspace paths per agent
- portkey — LLM routing through which all agents run
- supabase — Supabase CCP (svueekfvfrvhylxygktb) stores per-agent memory and tool_calls
- 1password — credential layer all agents consume
Open issues / TODOs
- AGENT-REGISTRY.md is 26 days stale — needs reconciliation for
stefania/stefania-asst/defaultretirement,_retired/full inventory, and agent count correction from 45 → 44 -
_retired/contains 16 subdirs; only 3 retirements documented in CLAUDE.md (stefania, stefania-asst, default from 2026-04-24). 13 others (codex, deal-command-center, debugger, trackcommander, angel-calls, betteracq, dealcmd, h2-calls, h3-calls, investorlift-angel, investorlift-h2, investorlift-h3, stefania-asst-retired-2026-04-24, stefania-retired-2026-04-24) are NOT documented in any retirement log. Requires Phase 1.5feedback_archive_not_deleteaudit. -
omni-contact-resolver.service(atlas-owned) is in❌ failedstate in AGENT-REGISTRY.md — per G-FAILED-SERVICE-MTTR, if >24h must be fixed, disabled, or archived - Wave 1 agent
_summary.mdstealth hubs: aurora (84 incoming), acquisitions (58 incoming), atlas (55 incoming), anthropic (50 incoming), dispo (41 incoming), backfill (38 incoming) — all flagged for W2-S0hub:false → hub:trueelevation. Not yet applied. - InvestorLift sub-agents (investorlift-angel, -h2, -h3) dir-count anomaly: these appear in BOTH
agents/(active) ANDagents/_retired/(archived). Duplication flag — investigate.
Recent activity
- 2026-05-03: hub created (W2-S3); live probes run; staleness documented
- 2026-04-24: stefania, stefania-asst, default retired per CLAUDE.md
- 2026-04-07: AGENT-REGISTRY.md last updated — reconciliation adding 9 agents