execution-graph

Execution Graph

Overview

The execution graph is the dynamic counterpart to the static workflow graph (Ch5). The workflow graph is the blueprint; the execution graph is the autobiography. Each query produces one execution graph instance — immutable after completion, queryable, the substrate for cognitive autopsy.

Each node represents one atomic operation:

• Node ID — unique per operation instance (span_id from OpenTelemetry in production)
• Node type — LLM_Call, Tool_Invocation, Retrieval, Decision_Point
• Timestamp — high-resolution for latency analysis
• Input payload — exact data received
• Output payload — exact data produced
• Performance metrics — latency_ms, token_count, cost_usd
• Parent-child edges — TRIGGERED relationships establishing causal lineage

Two-phase write is the central primitive (the chapter's execution-graph example):

1. Phase 1 (on-start): create node + link to parent. If the operation fails or crashes, the graph still captures what was about to happen.
2. Phase 2 (on-complete): fill in output + latency + cost.

A simple one-phase write loses the causal structure when nodes fail — exactly the cases that need diagnosis most.

When to Use

• BEFORE building any Ch7 evaluation layer (0/1/2/3) — they all query the execution graph
• Multi-step agent workflows where you need to attribute a failure to a specific node (the chapter quote: "the error isn't lost in a sea of model parameters")
• Self-evolution loops — every evolution decision needs the execution graph as input
• DevOps incident reconstruction — answer "which node failed and what preceded it"

Phrases: "execution graph", "trace the agent's run", "cognitive autopsy", "causal chain", "which node failed", "self-evolution foundation".

When NOT to Use

• One-shot single-call agents — no graph to trace; just log the call
• Production systems where you already pipe OpenTelemetry → Neo4j and have execution-graph reconstruction working
• High-frequency request paths where the per-node Cypher write latency blocks the request — buffer to async writer or sample