Agent Platform
Module:
operations:agents| Role: Any member (create), platform admin only (promote/demote) | Licence: Pro+ Tools:create_agent,update_agent,remove_agent,list_agents,view_agent,list_agent_tool_groups
The agent platform enables creation and deployment of specialized AI agents. Eve is the single orchestrator — users interact with Eve, who dispatches to specialized agents as needed. Agents never call each other directly.
Key Concepts
| Concept | Description |
|---|---|
| Agent | A specialized AI with a prompt, tools, and optional triggers |
| Agent Team | Organizational grouping of agents under a module with a shared mandate |
| Module | Commercial feature gate that controls which tools and agents are available |
| Dispatch | Eve calls an agent via a dispatch tool (e.g., ask_policy_researcher) |
| LANES | Design vocabulary for agent roles (see below) |
LANES Patterns
Every agent follows one of six patterns that determine its role and default behavior:
| Pattern | Role | Examples |
|---|---|---|
| Librarian | Store, index, retrieve knowledge | Policy Librarian (vault operations) |
| Analyst | Read sources, create briefings/reports | Policy Analyst (digests, summaries) |
| Navigator | Track work, flag risks, generate status | Policy Navigator (watchlist, deadlines) |
| Explorer | Monitor external signals, surface what matters | Policy Explorer (regulatory scanning) |
| Steward | Watch internal systems, detect deviations | Agent Steward (health monitoring) |
| Bespoke | Unique function not fitting other patterns | Custom agents |
LANES is not a class hierarchy — it provides prompt templates and tool defaults.
Agent Origins
| Origin | Created By | Modifiable | Visible To |
|---|---|---|---|
| Builtin | Code (engineering) | No | All workspaces with module |
| Workspace | Workspace members | Yes (by creator) | Current workspace only |
| Promoted | Admin promotion | Admin only | All workspaces with module |
Agent Modes
| Mode | Behavior |
|---|---|
| Execute | Takes action (creates scenarios, writes to vault, generates reports) |
| Track | Monitors and reports (watchlist updates, status digests, alerts) |
Scoping & Access Control
The platform uses a five-layer access model (enforced by check_agent_access()):
- Module gate — Is the required module enabled for this workspace?
- Workspace gate — Does the agent belong to this workspace (or is it builtin/promoted)?
- Origin gate — Is the agent's origin compatible with the operation?
- User visibility gate — For restricted agents, is the user in
visible_to? - Conversation scope gate — OAuth and signed-in sessions see private agents; shared connections see only shared agents
Visibility
workspace(default): All workspace members can see and interactrestricted: Only users listed invisible_tocan see and interact
Triggers
Agents can run autonomously on schedule or in response to events:
| Type | Description | Example |
|---|---|---|
| Cron | Scheduled execution | "0 6 * * *" (daily at 6 AM) |
| Event | Triggered by system events | Module enabled, data refresh |
| Webhook | External HTTP trigger | Partner API callback |
| Manual | User-initiated via Eve | "Run the regulatory scan" |
Delivery Channels
Triggered agents can push their output directly to the user who created the trigger:
| Channel | Description |
|---|---|
slack_dm | Direct message via the workspace's Slack bot |
email | Email to the user's registered address |
Configure per-trigger: "delivery": ["slack_dm", "email"]. See Creating Agents for details.
Related Documentation
- Creating Agents — How to create custom agents
- Agent Teams — Grouping agents into deployable modules
- Workflow Examples — End-to-end team deployment walkthroughs
- Policy Intelligence Team — Built-in policy agent team
- Module Licensing — What each module unlocks