Procedural Graph Runtime
Can a validation-gated, self-evolving execution graph improve long-horizon agent workflows without increasing unsafe tool use?
Hypothesis
Localizing the active graph node and translating its neighborhood into situational guidance beats flat memory or naive retrieval on completion and safety, at a comparable token cost.
Background
Seeded by Procedural Graphs (arXiv 2609.09153) and AgentGrad (arXiv 2609.08572) — the strongest architecture fit of the Sept 9 research run, though neither paper has a verified public repository yet.
Method
Build a small procedural-graph runtime; run the same multi-step workflow against a flat-memory baseline and the graph runtime; validate topology edits before committing them.
Measurement Plan
Task completion rate, wrong-order tool calls, latency, and token cost.
Expected Artifact
A graph runtime implementation plus a benchmark report comparing it to a flat-memory baseline.
Limitations
No public code was found for the source paper today — this starts as a clean-room prototype, not a reproduction.
Status
This experiment is Proposed — queued, not started. Results, data, and measurements will appear here once it actually runs.
