Workspace Capabilities
A structured guide to the workspace capability architecture: tools, execution paths, governance, workflows, web research, and sandbox analysis.
Workspace Capabilities
This guide maps the workspace capability architecture: what the system can do, which execution path owns each action, and how access is governed.
1. Capability Architecture
Workspace capabilities are organized into four layers:
- Policy Layer: Defines audience, risk, defaults, and aliases for each tool.
- Runtime Layer: Builds the available toolset for each agent and actor.
- Execution Layer: Runs tools in chat, workflows, or background phantom tasks.
- Governance Layer: Enforces owner/visitor boundaries, SQL safety, and write restrictions.
2. Owner Control Plane
Owner-only capabilities provide control over operations, settings, and review workflows:
- fetchInbox: Read pending/resolved inbox entries tied to the agent.
- resolveInboxItem: Mark inbox items as resolved with optional notes.
- listAvailableTools: View platform tools available for policy enablement.
- getAgentSettings / updateAgentSettings: Read and update behavior settings.
These tools are part of the owner control plane and are blocked for visitor sessions.
3. Workflow Execution Plane
Workflow capabilities let owners turn instructions into repeatable background execution:
- createWorkflow
- updateWorkflow
- fetchWorkflows
- triggerWorkflow
Workflows run through the shared orchestration path and can produce inbox alerts, summaries, spreadsheets, and documents.
4. Web Intelligence Plane
Web tools are aligned across interactive chat and background workflows:
- webSearch: Multi-provider cascade:
1. Firecrawl (primary)
2. Jina (fallback)
3. Legacy Supabase web-search fallback
- webScrape: Deep page reading with browser-first extraction and fallback fetch.
Use this plane when the agent needs current external information or must inspect a specific URL.
5. B2B Search & Prospecting Plane
B2B contact and company directory prospecting tools:
- b2bSearchPeople: Search for professional contacts, prospective leads, and people profiles using search queries, keywords, job titles, locations, company domains, and seniorities.
- b2bSearchOrganizations: Search for company profiles, organizations, and business structures using company name, domains, locations, employee ranges, and industries.
- b2bGetContactDetails: Retrieve full verified contact details, social links, and email/phone info for a contact by contact ID or email address.
6. Data and Sandbox Plane
Owner-scoped data tools are wired from the shared sandbox builder:
- runSQL: Read-only SQL querying (SELECT / WITH) with validation.
- executeCode: Restricted JavaScript execution for calculations and transformations.
- analyzeData: Read-only analytical SQL with automatic numeric summaries (min, max, avg, sum, count).
7. Shared Registry Design
The registry is designed to prevent drift between what policy exposes and what runtime can execute:
- Missing references were implemented or removed.
- The shared buildPhantomTools path now acts as a single source of truth for workflow tool execution.
- Sandbox tools are assigned as a complete group instead of partial/manual references.
8. Governance Boundaries
- Tool policy lives in the worker and cannot be bypassed from UI text.
- Owner-only tools remain inaccessible to visitors.
- Write/admin operations are policy-gated and role-checked.
- SQL pathways are restricted to non-mutating queries.
9. Recommended Operating Pattern
1. Use b2bSearchPeople + b2bSearchOrganizations to find B2B prospects and organizations natively.
2. Use webSearch + webScrape for external research requests.
3. Use runSQL + analyzeData for internal metrics and trend analysis.
4. Use createWorkflow / triggerWorkflow for repeated reporting tasks.
5. Route outcomes to inbox via reporting flows for owner review.
This pattern keeps execution auditable and consistent across interactive and background tasks.