Recommended first route
Your situation
I have a new project idea and want to explore it
Why this route
Exploration and prototyping is Scout's core loop, bounded cycles with visible output.
Worker map
⚡ Agent Capabilities
Use this page to pick the first worker, the first message, and the right handoff boundary fast.Use this page when your real question is who can do the work, what tools they have, and where their operating boundaries are.
Validated · Oct 6, 2026Revalidated on October 6 — Clawdy's model chip converged to GPT-6, verified by reading Clawdy's own gateway config over Bridge. The charter-family census pass converged the Scout Sub-Projects census onto the registry's five standing charters — Podcast Platform now cards beside Watchtower, Atlas, Forge, and Anvil as the newest standing loop. The Buzz card teaches the registry role (Budgets social loop). Current Scout capabilities use direct context and Linear-driven intake; advisor-bridge guidance is historical.Revalidated on October 6: Clawdy's model chip converged to GPT-6, verified by reading Clawdy's own gateway config over Bridge. The charter-family census pass converged the Scout Sub-Projects census onto the registry's five standing charters — Podcast Platform now cards beside Watchtower, Atlas, Forge, and Anvil. The Buzz card teaches the registry role (Budgets social loop). Current Scout capabilities use direct context and Linear-driven intake; advisor-bridge guidance is historical.
Take the fastest worker jump
Open First contact for a concrete situation, Intent guides for the likely owner, or Tie-breakers when two workers both fit.
Top-level route contracts
Pick the lane, then the page
Atlas stays easier to scan when the route map reads like navigation first. Start with the lane that matches your question, then expand the full page-job model only when you need the deeper contract details.
Current lane: Ownership
CapabilitiesHere
Worker-selection page for choosing the right agent and understanding execution boundaries.
Who owns the workCurrent pageSteady
Change pace: Usually current for a while, but revisit when routes, ownership, or workflows shift.
Best next move
Then route the work
After picking the worker, move to the handoff layer if you need to start the interaction well.
Open CoordinationStart here
Entry
Use these pages when you are still orienting and need the safest first click before you pick an owner or system.
3 pages
Who owns the work
Ownership
Use these pages when your real question is who should handle the work, how it moves, or which charter owns the surface.
4 pages
How the network is built
Systems
Use these pages when you need the map, structure, services, and interfaces behind the network instead of the people or routing rules.
5 pages
What to do when it hurts
Operations
Use these pages when something feels broken, you need a concrete next check, or you need the live-status companion surface.
2 pages
Reader lenses
Pick the routing layer that matches your question
Use these lenses when you do not want to read Capabilities top to bottom. Pick the lane that matches your question, keep the current focus summary in view, then jump directly to that section.
Short version: route by intent, switch to first-contact for real situations, use tie-breakers for close calls, then confirm with the full worker card.
Current focus
Match the exact moment to the right first message
Use the first-contact lane when you do not want theory first. It turns a real situation into the most likely first worker or route without making you scan every card.
Best for: Best for urgent routing, operator questions, and moments when you need to know who to contact before you care about the full capability model.
Stop when: Stop after First contact once you have one best first route, then drop into tie-breakers or a full agent card only if the boundary still feels fuzzy.
Intent Guides
If you already know the job to be done, pick the agent that owns that bottleneck first. Then use the full cards below to confirm tools, access, and boundaries.
ScoutCurrybotMegAtticusSamwiseGrafton
Improve one project or doc surface
Best for bounded research, docs/product improvements, and one-cycle builds inside a named charter or project.
Get shared engineering help
Best for team-facing implementation, PR review, fixes, and broader engineering support across repos.
Set design direction or audit UI quality
Best when the real question is layout, visual language, or whether an interface meets the bar before or after build work.
Make or unlock a decision
Best for authority, policy, and decision-boundary questions where the blocker is approval or governance, not execution.
Own the queue and move delivery forward
Best for triage, flow control, delivery pressure, and full task ownership from intake through completion.
Check Seed readiness or initiative health
Best for promotion readiness, gate status, lifecycle blockers, and network-level initiative health when the question is progress rather than implementation.
First Contact
The intent guides above work well when you know the category of work. This table handles the other case — you have a specific situation right now and need to know who to route it to.
- Recommended first routeYour situation
A PR needs review before merge
Why this routeTeam-facing code review and engineering support across repos.
- Recommended first routeYour situation
I want to add a new agent to the network
Why this routeScout owns Forge, the workspace generator that scaffolds new agent configs.
- Recommended first routeYour situation
I need a design decision before building UI
Why this routeMeg owns PDLP guidance and design direction, consult before, audit after.
- Recommended first routeYour situation
I'm not sure if an action is allowed by policy
Why this routeGovernance authority: classifies actions as auto-approve, needs-human, or deny.
- Recommended first routeYour situation
A new request needs to be routed into the right work item
Why this routeSamwise fronts intake on the Linear surface before work moves into execution.
- Recommended first routeYour situation
A task is stuck in the queue and needs triage
Why this routeSamwise owns WIP triage and priority routing across tasks.
- Recommended first routeYour situation
I need something executed end-to-end without hand-holding
Why this routeSamwise takes full task ownership from plan through delivery.
- Recommended first routeYour situation
I want to know if a Seed is ready for promotion
Why this routeGrafton runs hourly gate checks and owns the Seed → Work Item lifecycle.
- Recommended first routeYour situation
Something in the network feels broken or slow
Why this routeAtlas is educational, not a monitor. Start with named symptoms, then escalate.
- Recommended first routeYour situation
I want to check live agent health right now
Why this routeWatchtower is the real-time dashboard. Atlas documents how things work, not current status.
- Recommended first routeYour situation
I need context on a past decision or priority
Why this routeEarlier Scout cycles used the advisor bridge for historical context, priorities, and decisions; current cycles use direct context and Linear-driven intake.
- Recommended first routeYour situation
I want to understand the whole network at a glance
Why this routeThe network topology page shows all agents, systems, and connections in one interactive map.
| Your situation | Route to | Why this agent |
|---|---|---|
| I have a new project idea and want to explore it | 🔭 Scout | Exploration and prototyping is Scout's core loop, bounded cycles with visible output. |
| A PR needs review before merge | 🍛 Currybot | Team-facing code review and engineering support across repos. |
| I want to add a new agent to the network | 🔭 Scout | Scout owns Forge, the workspace generator that scaffolds new agent configs. |
| I need a design decision before building UI | 🎨 Meg | Meg owns PDLP guidance and design direction, consult before, audit after. |
| I'm not sure if an action is allowed by policy | ⚖️ Atticus | Governance authority: classifies actions as auto-approve, needs-human, or deny. |
| A new request needs to be routed into the right work item | 🕯️ Samwise | Samwise fronts intake on the Linear surface before work moves into execution. |
| A task is stuck in the queue and needs triage | 🕯️ Samwise | Samwise owns WIP triage and priority routing across tasks. |
| I need something executed end-to-end without hand-holding | 🕯️ Samwise | Samwise takes full task ownership from plan through delivery. |
| I want to know if a Seed is ready for promotion | 🌱 Grafton | Grafton runs hourly gate checks and owns the Seed → Work Item lifecycle. |
| Something in the network feels broken or slow | 🔧 (Troubleshooting) | Atlas is educational, not a monitor. Start with named symptoms, then escalate. |
| I want to check live agent health right now | 📊 (Watchtower) | Watchtower is the real-time dashboard. Atlas documents how things work, not current status. |
| I need context on a past decision or priority | 📋 Scout → Clawdy Advisor | Earlier Scout cycles used the advisor bridge for historical context, priorities, and decisions; current cycles use direct context and Linear-driven intake. |
| I want to understand the whole network at a glance | 🕸️ (Network page) | The network topology page shows all agents, systems, and connections in one interactive map. |
Tie-Breakers
Use these tie-breakers when the page-level roles feel adjacent. The fastest path is usually to choose the agent that owns the real bottleneck, not the one that merely touches the work.
Scout vs Currybot
You want code or repo help and both look plausible.
Choose this first
Start with Scout when the work is charter-owned exploration, a bounded prototype, or a docs/product improvement inside one project. Start with Currybot when the work is team-facing delivery, PR review, or shared engineering support.
Useful tie-breaker
If your first sentence sounds like "explore, prototype, map, or ship one cycle," pick Scout. If it sounds like "review, fix, notify the team, or help engineering broadly," pick Currybot.
Scout vs Atticus
The task involves a real decision, authority boundary, or policy question.
Choose this first
Start with Scout when you want someone to investigate, draft, or implement within known rules. Start with Atticus when the real blocker is "are we allowed to do this?" or "who has authority to approve this?"
Useful tie-breaker
Scout can prepare the case, but Atticus is the one that rules. If the next step needs a decision record instead of more execution, route to Atticus first.
Scout vs Meg
You are changing UI and need to know whether this is build work or design authority.
Choose this first
Start with Scout when you need the page or feature built. Start with Meg when the important question is layout, visual language, PDLP guidance, or whether a shipped UI meets the bar.
Useful tie-breaker
A good pattern is Meg before and after, Scout in the middle: Meg sets direction, Scout implements, Meg audits the result.
Scout vs Samwise
You need work done, but the ambiguity is whether this is a scoped build cycle or a delivery queue problem.
Choose this first
Start with Scout for project-specific research/build loops. Start with Samwise when the need is intake routing, triage, WIP control, or full ownership of a task from queue to completion.
Useful tie-breaker
Scout improves a named surface. Samwise manages flow and delivery pressure across tasks. If you are asking "what should move next?" or "who should own this from intake onward?" that is Samwise.
🧠
Clawdy
Primary Assistant & Network Hub
Model: GPT-6
Jake's main assistant and the network's orchestration hub. Runs as the host-level OpenClaw gateway — not a Docker container — with full delegation authority to all other agents.
Capabilities
💻
Full System Access
Shell execution, file operations, git, deployments
📱
Messaging
Signal, WhatsApp, Telegram, Slack, iMessage
🐙
GitHub Operations
PR reviews, issue management, CI/CD via gh CLI
🌐
Browser Control
Full browser automation via OpenClaw browser tool
🧠
Memory & Context
Semantic search across LC notes, team info, past conversations
🎛️
Agent Management
Can spawn sub-agents, query other agents via bridge
⏰
Cron & Scheduling
Create reminders, recurring tasks, automated check-ins
Example requests
Scheduling recurring work
“Remind me every Monday at 9am to review open PRs.”
System task with file access
“Run the deploy script in ~/scripts/deploy.sh and let me know if it errors.”
Complex reasoning request
“I have two options for the auth architecture — help me think through the tradeoffs before I decide.”
Access & Boundaries
Compute: Host-level OpenClaw gateway
Network: Tailscale mesh, internet
Storage: Full workspace read/write
Secrets: Full credential access
📋
Clawdy Advisor
Historical advisory pattern
Model: GLM-5.2
Read-only advisory pattern used by Scout for freshness and context recovery in earlier cycles. Keep this mental model when reading older workflow pages; current Scout cycles use direct context and Linear-driven intake instead.
Capabilities
❓
Context Queries
Answer questions about LC priorities, past work, decisions
🗺️
Network Topology
Provides current state of agent network, connections, roles
👁️
Read-Only Access
No actions, no side effects — purely advisory
Example requests
Network topology question (inside Scout cycles)
“What agents are currently active in the network? Have any new services been added since my last cycle?”
Priority check
“What is the current priority for Atlas? Is there anything Jake has flagged recently that I should know about?”
Access & Boundaries
Compute: Bridge only (via Clawdy session)
Network: Internal bridge calls only
Storage: Read-only via Clawdy's context
Secrets: None
🍛
Currybot
🔵 Domain TierLC Team Assistant
Model: GLM-5.2
Serves the entire Lambda Curry team. Handles PR reviews, coding tasks, team coordination, and shared workflows.
Capabilities
🔍
Code Review
Automated PR reviews, suggestion generation, lint/test feedback
⚙️
Coding Tasks
Feature implementation, refactoring, bug fixes via delegation
👥
Team Coordination
Slack notifications, standup summaries, team queries
🐙
GitHub Operations
Issue triage, PR management, CI workflows
🧠
Memory & Context
Access to LC team knowledge, project history
Example requests
PR review request
“Review PR #42 — focus on type safety and any missing edge cases.”
Bug fix
“The login form is failing on empty email. Fix it and add a test.”
Team notification
“Post to #engineering that the staging deploy is ready for QA.”
Access & Boundaries
Compute: Mac mini (host)
Network: Tailscale mesh, GitHub, Slack
Storage: LC team workspace
Secrets: Team-level credentials
🔭
Scout
🔵 Domain TierAutonomous Explorer & Builder
Model: GLM-5.3
Scout is the explorer-builder: researches, prototypes, and ships bounded improvements inside a named charter or project. Uses direct context and Linear-driven intake, delegates code writing when needed, and closes the loop with visible delivery updates.
Capabilities
🔄
Autonomous Work Cycles
Self-directed exploration with orient → work → log → report pattern
💪
Coding via Delegation
Delegates code changes to Codex (coding-muscle skill)
🌉
Current Context Intake
Direct context and Linear-driven intake for priorities and decisions
📦
Sandboxed Execution
Shell, git, bun, testing — all within Docker container
🌐
Browser Testing
Visual validation via agent-browser for UI work
📝
Memory & Logging
Progress logs, memory search, persistent state across cycles
📢
Slack Notifications
Failure alarms and human-needed escalations to #saffron-updates
Example requests
Research task
“Use Atlas search to find the route that explains how a handoff moves, then document which owner and surface answer the question.”
Build cycle
“Start an Atlas cycle — check freshness, ship one content improvement, and close with a Linear Project Update.”
Scoped prototype
“Prototype a dark-mode toggle in the nav. Do not open a PR — just get it working and screenshot it.”
Access & Boundaries
Compute: Docker sandbox (isolated)
Network: HTTP/HTTPS, bridge to other agents
Storage: Project workspace only (/home/node/.openclaw/workspace/projects/)
Secrets: Limited (API tokens via env vars)
Scout
Scout Sub-Projects
Scout keeps five standing charter loops on recurring cycles, each with its own workspace, progress log, and SCRATCHPAD — the registry's Scout product family: Watchtower, Atlas, Forge, Anvil, and Podcast Platform, the newest standing loop.
📊
Watchtower
Real-time observability dashboard for the agent network.
5175 · Meg mirror on 5181
🧭
Atlas
The educational map of the Saffron Network — this site.
18797 · Meg mirror on 18803
🔨
Forge
OpenClaw workspace designer and generator.
Pattern library + generator active
🔩
Anvil
AI rules research and audit engine.
Phase 1 active · research → ground → act cycle
🎙️
Podcast Platform
Multi-show podcast hosting and distribution built on the Saffron Starter stack.
Active · verified through private previews
⚖️
Atticus
🟡 Governance TierDecision & Governance Agent
Model: GLM-5.2
The governance layer for the agent network. Atticus does not do the build work. It decides whether proposed actions are allowed, what policy applies, and when a decision needs human authority or a durable ruling.
Capabilities
✅
Proposal Evaluation
Classifies proposed actions as auto-approve, needs-human, or deny
📋
Policy Enforcement
Maintains and applies governance rules without modifying individual agents
🔍
Knowledge Search
Queries QMD for precedent and established decisions before ruling
🔔
Human Escalation
Surfaces needs-human cases to Jake via Slack when input is required
📊
Audit Logging
Every decision emitted as an observability event (service: atticus)
🌉
Bridge Communication
Receives proposals and returns decisions via the Bridge protocol
Example requests
Governance check before acting
“I want to delete all logs older than 30 days from the observability store. Is that within policy?”
Policy question
“Does Scout have authority to merge its own PRs without human review?”
Access & Boundaries
Compute: Docker sandbox (isolated)
Network: HTTP/HTTPS, Bridge (receive-only proposals)
Storage: Read-only — no file modifications
Secrets: Observability token, QMD search token
🎨
Meg
🔵 Domain TierDesign Engineer
Model: GLM-5.2
Design authority for the Saffron Network. Meg sets the visual language, issues PDLP guidance, and audits shipped UI quality. Meg is the source of design direction, not the default implementer for every interface change.
Capabilities
📋
PDLP Generation
Creates Project Design Language Packages — versioned constraint files
📸
UI Screenshot Review
Evaluates UI output against PDLPs using vision models; returns punch lists
🧩
Storybook Component Library
Builds and maintains components in Storybook isolation
🔍
Visual Regression Testing
Playwright screenshot tests capture pixel-accurate baselines
🎨
Design Token Governance
Defines semantic color roles, type scale, and spacing tokens
🌉
Bridge Consultation
Available for design review requests via the Bridge protocol
Example requests
Design direction before building
“I am about to add a new table component to Atlas. Can you give me a PDLP snippet so it stays consistent?”
UI audit after shipping
“Here is a screenshot of the new onboarding page. Does it meet our visual standards? Any punch list items?”
Access & Boundaries
Compute: Mac mini (same host as LC containers)
Network: HTTP/HTTPS, Bridge (receive design review requests)
Storage: Per-project repo access for Storybook and PDLP files
Secrets: Observability token, QMD search token
🌱
Grafton
🟢 Operator TierLifecycle Manager
Model: GLM-5.2
Seed and initiative lifecycle manager for the Saffron Network. Named 'Grafton the Gardener', it monitors Seeds and promotes them to Work Items when readiness gate criteria are met.
Capabilities
🔄
Lifecycle Sweep
Hourly cron sweep evaluating Seeds for promotion readiness
🚦
Gate Evaluation
Evaluates each initiative against its current gate criteria
🔍
Blocker Identification
Surfaces blockers preventing promotion
📊
Initiative Health Tracking
Maintains awareness of initiative status across the network
📣
Lifecycle Summaries
Posts structured summaries to #saffron-updates after each sweep
🌉
Bridge Status Queries
Responds to Bridge queries requesting current initiative status
🌾
Harvest Loop
When a Work Item reaches Done, verifies result, wakes dormant Seeds, creates follow-ons
Example requests
Initiative status check
“What is the current gate status for the Forge initiative? Is anything blocking promotion?”
Harvest check after completion
“The Atlas PR just merged. Can you run the harvest loop and let me know if any follow-on Seeds should be created?”
Access & Boundaries
Compute: Mac mini (Docker container lc-grafton, port 18809)
Network: Bridge (receive status queries), QMD search
Storage: Own workspace memory, QMD remote search
Secrets: QMD search token, saffron-notify token
🧪
Dexter
🟢 Operator TierLabs Leader
Model: GLM-5.2
Lab leader for Lambda Curry Labs. Dexter guides experiment design, validation, and honest assessment of ideas so the founding team can test opportunities without pretending every idea deserves to become a product.
Capabilities
🧭
Experiment Design
Shapes candidate ideas into bounded experiments with a clear question, success signal, and next decision
📐
Validation Planning
Pushes teams toward falsifiable validation instead of build-first momentum
⚖️
Honest Assessment
Helps decide whether to continue, pause, or sunset an idea based on evidence
💬
Labs Coordination
Meets the team on direct surfaces like Slack, Notion, and CurryChat
Example requests
Choosing a lab experiment
“We have three product ideas. Which one deserves a two-week experiment first, and what would you validate before building?”
Designing a falsifiable test
“Turn this rough concept into an experiment with a clear hypothesis, success threshold, and shutdown condition.”
Access & Boundaries
Compute: Mac mini (chartered agent workspace)
Network: Slack, Notion, CurryChat, scheduled/manual triggers
Storage: Labs workspace and planning artifacts
Secrets: Scoped to Labs/operator-approved surfaces
📣
Buzz
🟢 Operator TierDistribution, Market-Signal & Budgets Social-Loop Agent
Model: GLM-5.2
LC Labs distribution and market-signal agent — Dexter's distribution counterpart. Buzz finds audiences, designs approved distribution experiments, sharpens messaging, turns learnings into playbooks, and owns the Budgets social loop (Social Operations + Social Studio) end to end.
Capabilities
🎯
Audience Discovery
Researches and identifies audiences for Labs experiments and product ideas
📊
Distribution Experiments
Designs bounded distribution experiments with human approval on public/outbound actions
✍️
Messaging Sharpening
Iterates on positioning and copy to improve reach and resonance
📱
Reddit Ads Management
Trusted read/write Reddit Ads management without per-action human approval
🗓️
Budgets Social Loop
Owner-executor of the Budgets social loop — strategy capture through scheduling human-Ready posts via social-ops publish tick
📋
Learning Playbooks
Turns distribution experiment results into reusable playbooks for the next cycle
Example requests
Audience research
“Who would actually use this? Can you research audiences and propose three distribution channels?”
Distribution experiment design
“Design a two-week distribution experiment for this feature with a clear success signal.”
Access & Boundaries
Compute: Mac mini (chartered agent workspace)
Network: Bridge, CurryChat, scheduled/manual triggers
Storage: Labs workspace and distribution experiment artifacts
Secrets: Scoped to Labs/operator-approved surfaces; Reddit Ads trusted carve-out
🕯️
Samwise
🟢 Operator TierDelivery Agent
Model: GLM-5.2
Single-owner delivery agent on Linear: Samwise shapes intake and queue judgment, keeps the shared conversation visible, and executes per-task work with its own skills; the retired Lantern and Spade charters remain as history.
Capabilities
🧭
Intake Routing
Samwise turns incoming requests on Linear into the right umbrella, owner, or next queue state
🏮
WIP Triage
Evaluates work in flight, decides priority, and routes tasks to the right executor
⛏️
Task Execution
Takes full ownership of individual tasks — plan, execute, verify, report
📊
WIP Limit Management
Prevents work from stalling in queues by tracking capacity
📣
Delivery Summaries
Reports task outcomes and delivery status through Linear at closeout
Example requests
Intake and routing request
“A new request just came in. Can you route it to the right work item or owner before it gets lost?”
Task triage request
“Can you triage the current WIP and let me know what should be picked up next?”
Execution status check
“What tasks are currently executing? Any blockers?”
Access & Boundaries
Compute: Mac mini (shared lc-builders gateway, port 18811)
Network: Bridge (receive tasks, coordinate with agents), QMD search
Storage: Own workspace memory, QMD remote search
Secrets: QMD search token, saffron-notify token
Skills
Skills System
Skills are reusable instruction sets that define HOW to use tools. When a task matches a skill's description, the agent reads the skill's SKILL.md and follows it.
🤖coding-agent🐙github🔒healthcheck🎨openai-image-gen🎤openai-whisper-api📜session-logs🛠️skill-creator🌤️weather💪coding-muscle