Progressive Crystallization: Turning Agent Exploration into Deterministic, Lower-Cost Workflows in Production

An evidence-gated lifecycle that progressively converts agent-discovered workflows into deterministic, zero-token playbooks reduced per-incident LLM inference cost by more than 70 percent while incident volume doubled, in a single production cloud-network operations deployment over eight months.

Observational production deployment at Microsoft Azure Networking: one agentic IT-operations platform handling tens of thousands of incidents per month, measured over eight months. The primary metric is the share of executions served by each of three execution types (agent-orchestrated, hybrid, deterministic) and the resulting per-incident inference cost trajectory. No randomized control group or holdout baseline is reported.

Deterministic (Type 1) executions rose from 0 percent to roughly 45 percent of volume; hybrid (Type 2) settled at roughly 30 percent; agent-orchestrated (Type 3) fell to roughly 25 percent. Per-incident agent cost fell by more than 70 percent over eight months while incident volume approximately doubled. Over 90 percent of common incident categories resolved autonomously; mean time to resolution fell from hours to minutes; false-positive remediation rate stayed under 5 percent. Token cost per run spans zero (Type 1) to 10k-50k (Type 3). Results hold only for a recurring-pattern-rich AIOps domain at one organization; the paper explicitly notes benefits shrink in environments dominated by novel one-off incidents.

Single organization, single operational domain: promotion thresholds and maturity ratios are not validated elsewhere. No controlled comparison: cost trajectory could partly reflect incident-mix shift, model-price drops, or other concurrent changes. Automatic acceptance-test generation from traces can produce under-specified tests, enabling premature promotion. Extraction quality depends on trace richness, limiting crystallization for poorly logged systems. The eight-month window captures one maturity curve and cannot show long-run steady-state cost or whether the Type 3 discovery pipeline remains effective as the pattern space matures. The paper is from a Microsoft employee describing a Microsoft Azure system, introducing potential reporting bias.

A fair skeptic notes this is a single-site case study with no control arm: the 70 percent cost reduction cannot be disentangled from concurrent LLM price drops, incident-mix changes, or other operational improvements over the same period. The promotion criteria (10 and 50 successful runs) are engineering choices, not empirically derived thresholds, and the acceptance tests are auto-generated from the same traces used to build the playbooks, creating circularity. The paper's author is affiliated with the deploying organization, and no independent replication or code release is described.

If the pattern generalizes, AIOps and similar repetitive-task agent deployments could bend their inference cost curves downward over time rather than scaling linearly with usage, which would meaningfully reduce the per-unit economics argument against broad agent deployment. However, the effect is domain-specific: it requires high incident recurrence rates. For the next 12 to 24 months the practical implication is modest: enterprises evaluating agent platforms for IT operations now have one production data point suggesting inference costs need not be treated as fixed per-incident, but the approach requires investment in trace infrastructure and lifecycle tooling that most organizations lack.

Market & Valuation: Enterprise software vendors and AIOps platform providers should assess whether this lifecycle pattern erodes their per-token or per-seat revenue as customers progressively route volume away from LLM inference.

Technical Capability: Researchers building agentic systems can treat progressive crystallization as a concrete alternative to fine-tuning or static workflow engines for amortizing agent discovery cost at scale.

Market & Valuation: Cloud infrastructure buyers operating at scale now have a production reference for negotiating LLM inference pricing commitments: volume growth need not equal proportional cost growth if crystallization is adopted.