I’m especially interested in feedback from people who’ve had to debug LLM-related production incidents or explain AI behavior after the fact.
Happy to answer questions or criticism.
0 comments
—
Show HN: IntentusNet – Deterministic Execution and Replay for AI Agent Systems · HackerTrans
I’ve been working on an open-source project called IntentusNet. It focuses on a narrow but persistent problem in AI systems:
AI executions are observable, but not reproducible.
When a production issue happens:
the model may already be upgraded
fallback logic may have changed
retries may be implicit
routing decisions are no longer recoverable
Logs tell you something happened, but they don’t let you replay the execution itself.
What IntentusNet does
IntentusNet is not a planner, prompt framework, or model wrapper.
It’s an execution runtime that enforces deterministic semantics around models:
explicit intent routing
deterministic fallback behavior
ordered agent execution
transport-agnostic agents (local, HTTP, ZeroMQ, WebSocket, MCP-style)
In the latest release, I added execution recording and deterministic replay.
Each intent execution can be:
recorded as an immutable artifact
replayed later without re-running models
explained even after models or agents change
The core invariant is simple:
The model may change. The execution must not.
Why I built this
Most AI systems implicitly trust the model to drive control flow. That makes failures hard to reason about and almost impossible to reproduce.
IntentusNet takes the opposite approach:
models are treated as unreliable but useful
routing and fallback are explicit and deterministic
executions are facts, not logs
This is closer to how distributed systems treat requests than how most LLM stacks work today.
Demo (what it actually proves)
There’s a small demo that shows:
A live execution with “model v1”
The same execution with “model v2” (different output)
A deterministic replay of the original execution, even after the model changes
Routing and execution order stay the same. Only the model behavior changes.
No debugger UI, no dashboards — just execution semantics.
What this is not
Not a replacement for MCP
Not a prompt-engineering framework
Not a monitoring system
Not trying to be “smart”
It’s infrastructure for making AI systems operable.
Repo
GitHub: https://github.com/Balchandar/intentusnet
I’m especially interested in feedback from people who’ve had to debug LLM-related production incidents or explain AI behavior after the fact. Happy to answer questions or criticism.