Overview
An agent in CelerFlow is a single AI agent instance connected to the control plane. Each agent has a unique identity, belongs to a workspace, and reports health and traces.
Agent states
CelerFlow tracks two independent status dimensions:
Connection status
| Status | Description |
|---|
connected | Agent is online and reporting. |
stale | No health report in 30+ days. |
disconnected | Gracefully disconnected by user. |
Termination status
| Status | Description |
|---|
active | Agent operates normally. |
terminated | All tool calls blocked. Set via kill switch. |
idle | Online but not actively processing. |
These are independent. An agent can be connected + terminated — it’s online but all tool calls are blocked until you re-enable it.
Agent identity
Each agent is identified by:
- Agent ID — UUID assigned on registration.
- Name — human-readable (defaults to hostname).
- Device info — hostname, OS, fingerprint, OpenClaw version.
- Bootstrap token — a rotating secret for authenticated API calls.
- Workspace path — the local directory the agent operates in.
Agent score
Every agent gets a 0–100 health score calculated hourly:
| Component | Weight | What it measures |
|---|
| Uptime | 30 pts | Last health check recency |
| Success rate | 25 pts | Tool call success vs error ratio (24h) |
| Block rate | 20 pts | Low = good; high block rate indicates misconfiguration or prompt injection |
| Response time | 15 pts | Average tool call duration (24h) |
| Config | 10 pts | Has active traces |
Kill switch
From the dashboard or API, you can terminate an agent instantly:
- Termination status is set to
terminated in the database.
- The CelerFlow plugin receives the update via Supabase Realtime (< 1 second).
- All subsequent tool calls are blocked with the message: “Agent terminated by user. Re-enable from CelerFlow Dashboard.”
- To restore, set the agent back to
active from the dashboard.
Termination is immediate. Any in-flight tool call will complete, but the next call will be blocked.
Risk discovery
CelerFlow automatically detects anomalies and surfaces risk alerts:
| Risk type | Trigger | Severity |
|---|
| Token spike | 1h token usage > 3× 7-day average | ⚠️ Warning |
| Latency spike | Avg tool call time > 5× 7-day average | ⚠️ Warning |
| High block rate | Block rate > 20% in the last hour | 🔴 Critical |
| Security anomaly | 3+ consecutive blocked tool attempts | 🔴 Critical |
| Offline | Health check overdue by 2× configured interval | ⚠️ Warning |
