AI Development as a Co-Evolving Stack

Alexander Polomodov treats AI development as a system of hardware, models, harnesses, tools, traces, and evals. The same model behaves differently with different context and authority: hardware sets economics, tools create consequences, traces support diagnosis, and evals grant release authority.

The layers move on two clocks: tool descriptions, routing, and release gates change in days; accelerators, serving, and training take months or years. Fix failures in the fast loop first; escalate only recurring errors from replayable episodes into post-training or infrastructure.

MCP standardizes tool calls but cannot ensure selection, safe authority, or an observed effect. An action contract covers identity, scope, idempotency or irreversibility, bounded output, legible errors, audit, and end state. Telemetry, eval material, and training data need separate retention rules; a trace does not prove an outcome.

For each episode, preserve start/end state, task contract, and layer versions; then change one layer and measure quality, variance, cost, safety, and human acceptance. Models and generic loops are rented; identity, policy, domain action contracts, and acceptance criteria stay owned. A custom harness needs a unique environment and evals.