Master scaling AI coding agents: Cut costs with best-of-N
Scaling the executor layer underneath your coding agent
Every agent session runs dozens of edits and test-fix cycles, and the executor layer underneath sets both the speed and the bill. In this session, Kwasi Ankomah shows how parallel execution, best-of-N selection, and disaggregated serving turn that cost into a controllable, predictable line item.
About this webinar
In Part 2 of the SambaNova coding agent series, Kwasi Ankomah moves from the agent harness itself to the infrastructure underneath it. The executor layer — the model that generates code, runs edits, and retries on failure — is the single biggest driver of both latency and cost in a coding agent pipeline. This session shows how to scale it without scaling the bill.
What you will learn
- Why the executor layer is the main cost and latency driver in a coding agent pipeline
- How to run multiple candidate solutions in parallel and pick the best one (best-of-N)
- How disaggregated serving separates prefill from decode to cut per-token cost
- How to measure and control agent spend as a predictable line item
- How these patterns apply regardless of which agent framework you use
Why executor scaling matters
A coding agent that works is only the first step — a coding agent that works affordably at scale is the real goal. Without deliberate control over the executor layer, costs grow linearly with every session, and latency compounds with every retry. The patterns in this session give teams a concrete path from “it works in a demo” to “it works in production at a price we can defend.”
Who should attend
AI and ML engineers building or operating coding agents, platform engineers responsible for inference infrastructure, engineering managers budgeting for AI-assisted development, and technical leads evaluating SambaNova or other inference providers. Following Part 1 of the series is helpful but not required.
Kwasi Ankomah is the Lead AI Architect at SambaNova Systems, where he builds production agentic AI systems on purpose-built RDU inference infrastructure. With 15 years of experience spanning financial services, consulting, government, and tech startups, Kwasi bridges deep technical expertise with a hands-on understanding of real business challenges. He specializes in deep agent architectures, multi-agent orchestration, and context engineering. Connect with Kwasi on LinkedIn →
