Scaling AI coding agents is where the real cost of automated software development gets decided. Kwasi Ankomah, Lead AI Architect at SambaNova Systems, joins Data Science Dojo to show how the executor layer sets both the speed and the bill.
About this webinar
This session is Part 2 of the SambaNova Sponsored Webinar Series, a collaboration between Data Science Dojo and SambaNova Systems built around the practical engineering choices behind agentic AI in production. Where earlier sessions introduced the building blocks of agent design, this one goes deeper into what happens once an agent starts writing and testing code at real-world scale.
Leading the discussion is Kwasi Ankomah, Lead AI Architect at SambaNova Systems, who spends his days building production agentic AI systems on SambaNova’s purpose-built RDU inference infrastructure. With 15 years of experience across financial services, consulting, government, and tech startups, Kwasi brings a rare mix of deep technical grounding and hands-on business judgment to multi-agent orchestration, deep agent architectures, and context engineering — the same patterns that power many of the coding assistants and open-source agent frameworks used every day.
What you will learn
Attendees will look closely at where a coding agent actually spends its time and money: not in the planning step, but in the executor, where dozens of edits, several rounds of test-fix cycles, and multiple candidate solutions run side by side. Kwasi will walk through how fanning that work across parallel SambaNova executors changes the economics of scaling AI coding agents, letting teams run more candidate solutions without a proportional jump in wall-clock time.
The session also covers best-of-N selection, where the test suite itself picks the winning candidate out of several parallel attempts, and why this approach depends on inference that is both fast and affordable at scale. Kwasi will connect this back to SambaNova’s disaggregated serving architecture, explaining the throughput, latency, and utilization tradeoffs that determine whether test-time compute is a cost center or a competitive advantage.
By the end of the session, attendees will understand the serving-level decisions that make scaling AI coding agents financially viable — and be able to evaluate their own agent infrastructure with a clearer sense of where the executor costs are actually coming from.
Why executor economics matters
As coding agents move from demos into daily engineering workflows, the cost of test-time compute has become one of the biggest hidden line items in AI infrastructure budgets. Running multiple candidate solutions per task multiplies inference calls, and without efficient serving, scaling AI coding agents can quietly turn into scaling AI coding costs instead. Understanding executor-level economics is what separates teams that can afford to let their agents explore multiple solutions from teams that have to ration every inference call.
This is also where infrastructure choices start to matter as much as model choice. Purpose-built inference hardware and serving strategies such as disaggregation are increasingly part of the conversation alongside model selection and prompt design. Readers who want more context on how DSD approaches applied AI engineering can explore the Data Science Dojo blog for related deep dives, and those curious about the inference platform behind this session can visit SambaNova directly.
Who should attend
This session is built for engineers and technical leaders who are already working with or evaluating coding agents in production — ML engineers, platform and infrastructure teams, and engineering managers responsible for the cost and reliability of AI-assisted development. It will also resonate with anyone scaling AI coding agents across a team or organization and looking to understand the trade-offs between speed, cost, and quality. A working familiarity with how coding agents and LLM inference operate is helpful, but no specialized SambaNova or RDU experience is required.