IntentusNet is a deterministic execution runtime for multi-agent AI systems. Make routing, fallback, and failure behavior replayable, explainable, and production-operable.
Multi-agent systems introduce failure modes that traditional software patterns don't address.
Agent selection varies between runs due to model drift, parameter changes, or race conditions.
Agents fail mid-execution with no record of partial state or clear recovery path.
Cannot reproduce failures or debug issues because outputs depend on live model calls.
Blocking one dangerous target blocks entire intents instead of allowing safe subset execution.
No structured logs, no execution traces, no way to audit what happened post-hoc.
Crash during multi-step execution leaves system in undefined state with no resume capability.
Explicit contracts for production multi-agent systems. No hand-waving, no "best effort".
Same input always produces same agent selection order. No implicit retries, no hidden state.
Execution state persisted before side effects. Recovery resumes from last safe checkpoint.
Every execution recorded with stable hashes. Replay returns recorded outputs without re-running models.
Block dangerous targets while allowing safe execution to continue. Filter, don't block wholesale.
Structured JSON output, grepable logs, SSH-friendly inspection. No dashboard required.
Every error categorized with typed error codes. No silent failures, no swallowed exceptions.
IntentusNet does not build agents, define prompts, call LLMs, or orchestrate workflows. It is a runtime layer that sits beneath agent frameworks to make their execution deterministic and debuggable.
Think of it as systemd for AI agents: it doesn't write your service, but it ensures reliable execution, restart semantics, and observability.
A deterministic execution pipeline from intent to completion.
Declare what to do
Deterministic agent selection
Filter dangerous targets
Record before effects
WAL-backed state
Replay & debug
$ intentusnet run --intent "power-off-for-maintenance"
{
"execution_id": "exec-a7b3c9d2",
"status": "completed",
"route": {
"strategy": "FALLBACK",
"agents_tried": ["hvac-controller", "lighting-controller"],
"selected": "hvac-controller"
},
"policy": {
"filtered": ["cctv-controller"],
"reason": "security_exclusion"
},
"replayable": true
}IntentA declarative unit of work (name + version + payload)
RouteThe deterministic selection of which agent handles an intent
PolicyRules that allow/deny/filter intent execution targets
Execution IDUnique identifier for each execution instance
WALWrite-Ahead Log for crash-safe state persistence
ReplayReturn recorded output without re-executing models
Start with a simple pip install. No configuration required for basic usage.
pip install intentusnet