Multi-Agent: Get the Most Out of One Agent First

In the second AMA, Alexander Polomodov and Aleksey Litvinov connect local models, multi-agent work, and legacy systems with security, verification cost, and a product’s ability to turn faster development into value.

A harness can offset a weaker local model, but needs hardware and maintenance. Quality criteria survive model replacement. Multiple agents provide independent roles and parallel search, yet reload context and spend tokens; start with one agent and add a role only for a measured constraint.

For an undocumented service, first capture current behavior, place characterization tests around one change, and build a short loop from compilation, linting, and tests. Technology cannot replace ownership of purpose. Security likewise begins with a threat model for the harness, model, tools, and data; controls should mitigate a concrete risk.

AI budgets work better by project or task type, traced through a corporate gateway. Local evals route repeatable work to a cheaper model and hard cases to a stronger one. Tokens and pull requests do not prove value: acceleration only moves the bottleneck. Outcomes belong in product behavior, customer response, quality, and end-to-end time.