| Hi folks — I’ve been working on IntentusNet, a small execution runtime that focuses on deterministic, replayable execution semantics around AI tools. The problem I kept hitting in production systems:
AI pipelines are observable, but not reproducible. After an incident, models, routing logic, retries, or fallbacks may have changed — logs alone don’t let you replay what actually happened. v1.3.0 introduces a runtime determinism core: Write-ahead log (append-only JSONL) written before side effects Crash-safe recovery and deterministic replay (fails loud on divergence) Runtime execution contracts (timeouts, retries, cost ceilings) Side-effect classification to prevent unsafe retries or fallback CLI-first inspection (list / show / trace / replay / diff) It’s not a planner or agent framework, and not a replacement for MCP — it focuses purely on execution semantics around tools (including MCP-style tools). Quick try (run from repo root): git clone https://github.com/Balchandar/intentusnet cd intentusnet
pip install -e .
python -m examples.deterministic_routing_demo.demo --mode with
python -m examples.deterministic_routing_demo.demo --mode mcp Docs (architecture, guarantees, demos):
https://intentusnet.com MIT licensed, open source:
https://github.com/Balchandar/intentusnet I’d really value feedback from people building real systems: What guarantees do you expect from deterministic replay in practice? How do you handle retries and side effects safely in AI pipelines? |