When your AI agents fail,
can you tell why?
CLowl gives every message a type, every task a state, every failure a trace.
Semantic layer for LLMs. State machine for runtimes. Observability for teams.
{
"p": "ERR",
"from": "audit-bot",
"to": "deploy-node",
"task": "audit",
"data": { "code": "E007", "msg": "Disk full", "retry": false },
"ctx": "sys-logs/disk-alert"
}
→ audit-bot tells deploy-node: ERROR on audit
"Disk full" (E007, not retryable)Semantic Layer
For LLMs
180-token system prompt injection. Agents declare intent in typed JSON. The LLM handles meaning. The runtime handles everything else.
Coordination Protocol
For Runtimes
9 states. 11 events. No backward transitions. Deterministic lifecycle tracking from CREATED to COMPLETED. Your state machine, not the LLM's guess.
Observability Platform
For Teams
Trace trees across agent chains. Cost-per-branch analytics. Failure lineage mapping. SLA tracking per state. Like OpenTelemetry, but for AI agents.
Three layers. Clean separation.
The LLM generates intent. The runtime enforces guarantees. The transport delivers messages. They never cross.
What the agent means. Performative, task, data, context reference. ~150 tokens. Stable. Strict schema. No hallucinated fields.
Message IDs, timestamps, trace propagation, auth, lifecycle state. Deterministic. Idempotent. The LLM never touches this layer.
MCP, A2A, HTTP, WebSocket, Kafka, gRPC. CLowl rides on top of any transport. It defines the payload, not the pipe.
Ten performatives. That's the whole language.
Every CLowl message has exactly one p field. These are the ten valid values.
| Code | Name | Meaning |
|---|---|---|
REQ | Request | Do this thing. Initiates a task. |
INF | Inform | Here is information. No action expected. |
ACK | Acknowledge | Got it, proceeding. |
ERR | Error | Failed. Here's why. Structured error code. |
DLGT | Delegate | Passing to someone better suited. |
DONE | Complete | Finished. Here is the result. |
CNCL | Cancel | Abort this task. |
QRY | Query | What's the status? Give me info without acting. |
PROG | Progress | Here's an update on the running task. |
CAPS | Capabilities | Here's what I can do. Broadcast on connect. |
Built for builders. Ready for enterprise.
Developers
Drop in a 180-token system prompt. Your agents stop inventing their own message formats. Works with GPT-4, Claude, Gemini, Llama, Mistral.
Architects
JSON Schema validation on every message. Formal state machine with deterministic transitions. Transport-agnostic. Version-safe.
Enterprise Teams
SLA tracking per lifecycle state. Audit trails for every agent handoff. Deterministic replay for incident review. Governance-ready.
See it in action
Paste any CLowl message. Get plain English. Or describe what you want and get valid CLowl JSON back.