Your AI Agent Forgets Everything It Learns: Here's Why That's Expensive
The knowledge loss problem. Every execution re-invokes full reasoning for problems already solved. The same 8-step pattern repeated 47 times, costing 30,000 tokens each time.
Progressive Crystallization
AI agents are deployed as permanent cost centers, re-invoking full LLM inference even for solved problems. This 10-part series explores how to systematically convert agent discoveries into deterministic, zero-cost automation that becomes cheaper, faster, and safer over time.
Paper published · posts coming soon
-
01
-
02
The Three Types of AI Automation You Need to Know About
Execution type taxonomy: Deterministic (Type 1, zero tokens), Hybrid (Type 2, structured + LLM), Agent-Orchestrated (Type 3, full reasoning). A spectrum from zero-cost to full-agent.
-
03
How to Turn an Agent's Best Work into Permanent Automation
Trace extraction algorithm, template generation, parameterization, and DAG construction. Converting successful agent runs into reusable deterministic workflows.
-
04
The Promotion Ladder: From Stochastic to Deterministic in 4 Stages
Evidence-based advancement through the crystallization lifecycle. Promotion requires 90%+ success rate, less than 5% false positives, and zero rollbacks for 30 consecutive days.
-
05
Zero Tokens, Zero Problems: The Economics of Workflow Crystallization
A crystallized playbook runs in 200ms at $0.00. The same resolution via full agent reasoning takes 45 seconds and costs $0.35. Token costs dropped 72% over 8 months.
-
06
Why Safer AI Gets Cheaper: The Safety-Determinism Connection
Every safety property improves or is maintained across each promotion stage. Deterministic playbooks can be formally verified; agent reasoning cannot.
-
07
Process Mining Meets LLMs: Extracting Workflows from Agent Traces
Applying process mining foundations to agent execution traces. Each execution produces structured logs: ordered tool invocations with typed I/O, decision branches, and timing.
-
08
When Deterministic Fails: Graceful Demotion and the Discovery Pipeline
Circuit breakers, fallback routing, and continuous discovery. How to handle when a crystallized playbook breaks and needs to fall back to agent reasoning.
-
09
From Runbooks to Crystallized Playbooks: AIOps in Practice
Enterprise case study with 90%+ autonomous resolution rate. Resolution time decreased from hours to minutes by crystallizing agent discoveries into deterministic workflows.
-
10
The Maturity Metric: Measuring Your Platform's Crystallization Progress
The Type 1:2:3 execution ratio as a platform health metric. Tracking the shift from 100% agent to 45% deterministic, 30% hybrid, 25% agent over 8 months.
This series unpacks the paper "Progressive Crystallization" (arXiv:2607.07052), grounded in production experience building an agentic AIOps platform at Microsoft Azure Networking. The paper is out now; the posts below break it down one idea at a time. First post coming soon.