Agents
Cantila's self-healing brain runs an agent swarm over your projects on a 60-second tick. High-confidence, safe actions auto-apply; destructive ones wait for your confirmation. Every action is logged, and a learning loop keeps the system honest over time.
The swarm
Eight specialized agents run on each tick:
| Agent | Watches |
|---|---|
| Uptime | Health checks, restarts, rollbacks |
| Deploy | Build / release health |
| Cost | Spend anomalies and waste |
| Scale | Right-sizing replicas |
| Security | Exposure and credential hygiene |
| Capacity | Headroom on shared resources |
| Sending health / deliverability | |
| SMS | Telephony health |
What auto-applies vs. waits
- Auto-apply (high-confidence, safe): rollback a bad deploy, scale down idle capacity, sleep an idle project.
- Waits for confirmation: anything destructive surfaces as a proposal you approve.
Every action — auto-applied or proposed — appears in the activity feed at
/v1/activity.
Trust over time
A learning loop downgrades confidence for action kinds with a bad track record, so a class of fix that has caused problems before stops auto-applying. Post-checks verify auto-applied actions and flag regressions.
Pausing
Pause the swarm at any time:
cantila agents pauseResume when you're ready:
cantila agents resumeAPI reference
| Method | Path | Notes |
|---|---|---|
GET | /v1/agents/status | Current swarm status |
POST | /v1/agents/pause | Pause auto-apply |
POST | /v1/agents/resume | Resume |
GET | /v1/agents/proposals | Pending proposals to confirm |
GET | /v1/activity | Full action log |
MCP & Console
- MCP:
cantila_agents_status - Console:
/agents