Year: 2026
Type: dissertation
Abstract: Coding agents complete software tasks autonomously over many steps, yet the record they leave behind is incomplete: a finished session preserves its outcome, while the process that produced it survives only as an unstructured event log that can be reread but not queried. This thesis introduces the Interaction Graph, a typed provenance model for human-agent software workflows lifted from a raw OpenTelemetry span stream, extending the W3C PROV standard for language-model agent runs. It is evaluated under a Design Science methodology through three research questions, each against objectives fixed in advance. The first question concerns representation. The lift reconstructs the version-controlled history and classifies each action at a level of agreement with human annotators sufficient to certify the graph as a reliable substrate. The representation strictly dominates the flat evaluation log in query expressiveness, answering every question the flat log admits plus a further set answerable only after the lift, one a language model recovers from the flat log alone only in part. The second question concerns diagnosis. Recurring struggle motifs are expressed as deterministic subgraph queries, validated against a human-annotated gold standard in which a fired detector is almost never a false alarm. Motifs' association with failure stays statistically indistinguishable from chance at this sample size. The queries nonetheless execute at a fraction of the cost of a per-trace language-model judge. The third question concerns memory. A null-tolerant feasibility pilot integrates the graph as a cross-session memory a coding agent queries during a task. Task resolution is unchanged, as expected of a process-oriented memory, while tool-call repetition is measurably reduced, though one configuration also shows a regression on a failure-predicting motif, visible only in the graph's typed layers. A single typed representation, lifted once from the event stream, thus serves as a query substrate, a diagnostic surface and a cross-session memory without modification. This unification constitutes the central contribution of the thesis.
Source: Aaltodoc (Aalto University)
Author: Sokhin, Artem
Language: English
Subfield: Information Systems and Management
Field: Decision Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7204021391