Conflict as Telemetry for Illegible AI: Governing LLM Agent Workflows (Extended Abstract)

Abstract

Modern enterprises are moving from generative AI as a productivity layer to autonomous large language model (LLM) agents that execute multi-step workflows with limited human oversight. This shift amplifies an under-instrumented management risk: organizational illegibility, where artifact polish and apparent operational smoothness decouple from understanding, accountability, and recoverability. When decisions cannot be reconstructed, governance becomes performative, auditability becomes brittle, and incident response degrades into blame allocation instead of safe reversal. We propose a runnable field protocol that treats recurring conflict (repeated escalations, reversals, exception requests, rework loops, blame cycles) around mission-critical workflows as early-warning telemetry for illegibility. The method couples five illegibility fractures (Decision Fog, Knowledge Drift, Synthetic Competence, Ghost Apprenticeship, Promotion Blindness) with six recurrent behavioral patterns under load (Cynic, Ghost, Hero-Martyr, Passive Burnout, Underminer, False Compliance). Teams score fracture-pattern pairs, select the top two, then run the loop. The coupling is used diagnostically: conflict is not treated as a morale defect, but as a signal that decision rights, knowledge distribution, and risk ownership have drifted out of alignment under automation pressure. Patterns are not traits; they are roles produced by misaligned accountability. This is not a character audit. It is a legibility audit. Field Procedure: Map–Probe–Trace–Teach (MPTT) • Inputs: Active conflict, agent workflow, ownership map • Steps: (1) Score fracture×pattern pairs, select top 2; (2) Map decision chain end-to-end; (3) Probe for gaps in explanation/ ownership; (4) Trace lineage to specific owners/ artifacts; (5) Teach via runbooks, tests, review gates • Outputs: Owner map, decision record, rollback test, gate update • Stop-ship: Time-to-explain unchanged after 2 cycles, rescope The illegibility fractures manifest predictably in agentmediated environments. Decision Fog appears when operators can observe outcomes but cannot reconstruct the reasoning or execution chain that produced them. Knowledge Drift appears when operational expertise migrates into prompts, tooling glue, or model behavior while documentation, training, and review practices lag behind. Synthetic Competence appears when agent-assisted throughput is mistaken for organizational capability, leaving the enterprise unable to debug, modify, or safely extend workflows when conditions change. Ghost Apprenticeship appears when learning is mediated primarily by agent outputs rather than by senior practitioners, producing tool-shaped judgment without durable mental models. Promotion Blindness appears when advancement tracks visible output velocity while discounting independent capability, creating accountability voids when the agent encounters novel conditions. For a selected agent-mediated workflow, teams run a fourverb loop: Map the end-to-end decision chain across humans, tools, and models; Probe for missing explanations and unverifiable claims; Trace responsibility, data lineage, and change history to specific owners and artifacts; Teach by updating runbooks, tests, review gates, and escalation rules so the organization can reproduce the reasoning under pressure. Each cycle produces four artifacts: owner map, decision record, rollback test, gate update. We define measurable outcomes suitable for enterprise evaluation: shorter time-to-explain and time-to-reverse during incidents, increased completeness of decision records, reduced recurrence of the same conflict around the same workflow, and fewer unowned risk transfers between teams. If time-to-explain does not decrease by 30% after two Map–Probe–Trace–Teach cycles, the protocol has failed for this workflow and requires re-scoping. The protocol is most effective when conflict recurs around workflows where agents trigger irreversible actions, multiple teams share ownership of agent-mediated processes, or regulatory requirements demand demonstrable human oversight. The method is least effective when conflict stems from resource scarcity, interpersonal dynamics unrelated to work structure, or external constraints no internal governance change can address. These boundary conditions make the approach falsifiable in practice. The contribution is not a new model or agent architecture, but a falsifiable management instrument for intelligent transformation that raises reconstructability as agent autonomy scales. The method is intended for deployment contexts where agent decisions can trigger costly or irreversible actions (customer commitments, financial transactions, production changes) and where audit trails must demonstrate that human judgment and accountability remained legible.