Agent Capabilities

Cortex provides 56 granular capabilities across 13 categories, enabling fine-grained control over what each agent can do.

Capability Categories

Agent Management (5 capabilities)

Capability	Description	Dangerous
agent.create	Create new agent definitions	No
agent.read	View agent configurations	No
agent.update	Modify agent settings	No
agent.delete	Remove agents	Yes
agent.execute	Trigger agent execution	No

Task Management (5 capabilities)

Capability	Description	Dangerous
task.create	Create kanban tasks	No
task.read	View tasks	No
task.update	Modify task status/details	No
task.delete	Remove tasks	No
task.assign	Assign tasks to agents	No

Skill Management (4 capabilities)

Capability	Description	Dangerous
skill.create	Create new skills	No
skill.read	View skills	No
skill.update	Modify skills	No
skill.delete	Remove skills	No

Swarm Coordination (5 capabilities)

Capability	Description	Dangerous
swarm.create	Create agent swarms	No
swarm.read	View swarm status/metrics	No
swarm.update	Modify swarm configuration	No
swarm.delete	Remove swarms	Yes
swarm.execute	Launch swarm execution	No

Integration Management (4 capabilities)

Capability	Description	Dangerous
integration.create	Add new integrations	No
integration.read	View integration configs	No
integration.update	Modify integrations	No
integration.delete	Remove integrations	Yes

Webhook Management (4 capabilities)

Capability	Description	Dangerous
webhook.create	Register webhooks	No
webhook.read	View webhook configs	No
webhook.update	Modify webhooks	No
webhook.delete	Remove webhooks	No

Memory & Knowledge (4 capabilities)

Capability	Description	Dangerous
memory.create	Store memory entries	No
memory.read	Query/search memories	No
memory.update	Modify memory entries	No
memory.delete	Remove memory entries	No

Document Management (4 capabilities)

Capability	Description	Dangerous
document.create	Create documents	No
document.read	View documents	No
document.update	Modify documents	No
document.delete	Remove documents	No

LLM Provider Management (4 capabilities)

Capability	Description	Dangerous
llm.create	Add LLM providers	No
llm.read	View provider configs	No
llm.update	Modify providers	No
llm.delete	Remove providers	Yes

Organization Management (4 capabilities)

Capability	Description	Dangerous
org.read	View org settings	No
org.update	Modify org settings	Yes
org.members	Manage org members	Yes
org.billing	Access billing data	Yes

Scheduling (4 capabilities)

Capability	Description	Dangerous
schedule.create	Create cron jobs	No
schedule.read	View schedules	No
schedule.update	Modify schedules	No
schedule.delete	Remove schedules	No

Monitoring (5 capabilities)

Capability	Description	Dangerous
status.report	Report agent status	No
status.read	View status reports	No
metrics.read	View agent metrics	No
audit.read	View audit logs	No
health.read	View system health	No

Communication (4 capabilities)

Capability	Description	Dangerous
message.send	Send inter-agent messages	No
message.read	Read messages	No
notification.send	Send notifications	No
notification.read	Read notifications	No

Dangerous Capabilities

These 10 capabilities require elevated access and may trigger confirmation flows when using write-level access:

1. agent.delete — Permanently removes an agent definition
2. swarm.delete — Destroys a swarm and all its associations
3. integration.delete — Removes integration and its credentials
4. llm.delete — Removes LLM provider config (may break dependent agents)
5. org.update — Changes organization-level settings
6. org.members — Can add/remove organization members
7. org.billing — Access to billing and payment data
8. system.admin — Platform-level admin operations
9. system.config — Modify system configuration
10. data.export — Export bulk data from the platform

Preset Profiles

Seven built-in profiles provide sensible defaults:

Profile	Use Case	Enabled Capabilities
Chief	Strategic coordination (never writes code)	All read + task/swarm management + monitoring
Coder	Implementation + self-testing	Agent/task/skill/memory CRUD + execution
Tester	QA validation	Read all + task update + status report
Researcher	Investigation + analysis	Read all + memory/document CRUD
Architect	System design	Read all + task/document CRUD
PM	Project management	Task/swarm CRUD + monitoring + org read
Viewer	Read-only observer	All read capabilities only

Capability Resolution

Effective capabilities are resolved through hierarchical intersection:

effective = definition_caps ∩ instance_caps ∩ project_caps ∩ workspace_caps

Each level can only restrict — never grant capabilities that a higher level denies.

Resolution API

curl -X POST https://api.cortex.acrobi.com/api/agent-capabilities/resolve \
  -H "Authorization: Bearer $TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agentName": "my-coder-agent",
    "scope": "project",
    "entityId": "proj-123"
  }'

Response:

{
  "capabilities": {
    "agent.create": true,
    "agent.read": true,
    "agent.delete": false,
    "task.create": true,
    ...
  },
  "profile": "coder",
  "scope": "project",
  "resolvedAt": "2026-04-03T12:00:00.000Z"
}

See Capability Control for the 4-tier access level system.