ACM Projects Three Agent Frameworks into Versioned Configuration Graphs, Unifying Change-Impact and Runtime-Provenance Tracking

The behavior of agent systems is typically distributed across prompts, model names, tool permissions, skill files, handoff relationships, and framework-specific workflows. Even when the code is under version control, teams still struggle to determine which agents will be affected by a model or prompt change, which configuration was actually used for a given run, and whether different frameworks follow the same governance rules.

A new technical report proposes Agentic Configuration Management (ACM), which represents these components as typed, independently versionable Agentic Configuration Items. Each item retains an immutable revision and is assembled into a baseline at deployment time. Native framework configurations are mapped through semantic projection into a canonical Configuration Graph. The graph explicitly records dependencies, composition, tool usage, and agent handoff relationships. Configuration versions are also separated from runtime provenance, preventing execution records from containing only framework-internal objects or text snapshots that cannot be reconstructed.

Change-impact analysis uses monotonic propagation over a finite lattice. Starting from the modified nodes, it repeatedly computes a fixed point based on relationship and change types until the affected set stops growing. The authors provide a Python scaffold and adapters for LangGraph, CrewAI, and the OpenAI Agents SDK. The paper evaluates the approach using 27 governance scenarios and nine cross-framework cases. In one example containing 13 configuration items and 18 relationships, a single-hop dependency check identified only six items, while fixed-point propagation identified the same nine found through a complete manual trace. Each of the nine cases was rerun five times, with deterministic results.

For teams that need to audit multi-agent deployments, ACM's value lies in establishing a configuration supply chain above orchestration frameworks, rather than creating another agent runtime. It can turn questions such as “What will be affected if we change the model?” into graph queries, while providing a shared data model for rollbacks, approvals, and runtime-provenance tracing.

However, the current evidence remains at the reference-implementation level. The oracles for the nine cases were predefined by the authors rather than independently assessed by a third party, and the equivalent dependency structures across the three frameworks were deliberately configured. Some dynamic CrewAI topologies cannot be introspected statically and still require supplemental adapter metadata. The public repository is also at an early stage and does not demonstrate reliability in large-scale production environments, under malicious configurations, with dynamically generated tools, or amid high-frequency version changes. Engineers should therefore treat ACM as an experimental schema and governance model, rather than a mature standard.