Procedural Graphs: Teaching Agents What To Do Next
This presentation explains how Procedural Graphs give language-model agents explicit procedural memory by representing transitions between actions as an editable, self-evolving graph structure. The method combines localized retrieval of valid next steps with iterative refinement from execution feedback, improving long-horizon performance on financial simulation, function calling, and multi-step reasoning tasks without hard-coding rigid workflows.Script
Language model agents usually keep procedural knowledge locked inside their parameters or buried in prompt text. When an agent must decide whether to forecast cash flow before requesting financing or check prerequisites before calling a tool, that logic stays invisible and uneditable.
Procedural Graphs externalize that structure as an attributed directed graph. Each edge states that one procedure can follow another under a specific condition, includes recommended guidance for that transition, and warns about execution hazards like submitting duplicate requests.
At inference time, the system matches the agent's current action to a graph node, retrieves its local neighborhood up to two hops away, asks a guidance model to verbalize that subgraph into situational advice, and lets the solver choose its own next step. This localized soft control outperforms injecting the full graph by nearly 30 points on embodied tasks while cutting token use by two thirds.
The graph evolves through an offline refinement loop that executes training tasks, proposes mutations like adding nodes or revising edge conditions, validates candidates on held-out episodes, and accepts only those that match or exceed the current graph's validation score. On a long-horizon financial task, ten rounds of evolution raised survival from zero percent to 85 percent while reducing monthly tool calls by more than 80 percent.
Across six benchmarks and four language models, Procedural Graphs ranked first or tied for first in 21 of 24 combinations, with the largest gains appearing on function calling, professional workflows, and policy-constrained interaction. The method improves not by reducing tool use universally, but by reordering actions so that prerequisite checks and forecasts happen before critical decisions.
Procedural knowledge can be learned from execution and refined when expert initialization fails. If you want to explore how self-evolving execution structures might reshape agent design, visit EmergentMind.com to dive deeper into this paper and create your own research videos.
