This is the first DevTalks webinar hosted by SambaNova. Justin Woo, developer advocate at SambaNova, is joined by Shane from CrewAI to walk through how effective AI agents get designed, then build a working multi-agent crew live from a single natural-language prompt. The session covers how AI applications evolved from chatbots to agentic workflows, how to choose a model per agent, and why inference speed and cost matter more for agents than for chat.
TL;DR
- Agentic workflows use multiple specialized agents that mirror real human roles, rather than one general-purpose assistant. CrewAI generated four agents from one prompt in the live demo: market research specialist, product analysis expert, customer sentiment researcher, and product evaluation analyst.
- Prompt detail is the main lever on quality. A few sentences is enough to get started, but the more specific the instructions, the closer the output gets to a production system.
- Agents do not all need the same model. You can swap models per agent and replay a single task to compare. If a smaller model gets you 95% of the way, it is usually the right choice.
- Agents are token-intensive because they reason, plan, explore and verify, so a single request can trigger many LLM and tool calls. That makes fast, cost-efficient inference more important for agents than for a chat interface.
How AI applications evolved into agentic workflows
Shane frames the shift in three phases, which is a useful way to place your own systems.
Phase one: chatbots
A single, general-purpose agent trained on what is available on the internet. You ask questions and it generates answers. Narrow in scope, and in most workplaces it arrived as a broad mandate to start using AI for things like writing an email.
Phase two: assistants and copilots
Almost all built with RAG, because enterprises concluded the models needed access to their own domain and business knowledge to be useful internally. Typically one agent, sometimes two or three, and highly domain specific, which is why most companies had to build their own. The goal was employee productivity, such as extracting information from long financial statements or answering questions about patient records.
Phase three: agentic workflows
Multiple agents, sometimes many. The idea is to build specialized agents that reflect the real world: if a marketing team has someone doing demand generation, someone managing ads and someone on product marketing, the agents mirror those roles. They target repeatable processes involving a lot of tools, people and steps, and deliver the most value when they are autonomous, taking action rather than waiting to be asked.
Shane makes the point that orchestration is foundational rather than sufficient. The value comes from being able to build, observe, optimize and scale agents, which is where visual editors, tracing, guardrails and human-in-the-loop feedback come in. A General Assembly case study in the session shows agents mapped one-to-one onto real team roles, including an instructional architect and a QA person, cutting lesson development time by 90% and reducing team burnout.
Building a crew from a single prompt
The live demo starts with a prompt describing a workflow: identify the top three products in a category, research them, search the internet, compare them, and compile a report. CrewAI’s visual studio reasons through it in real time and generates the full workflow, defining the agent roles, the tasks and the backstories that serve as each agent’s system prompt.
Justin points out that auto-generating the backstory removes real cognitive load, since writing a good system prompt by hand is one of the harder parts of building agents. Both presenters treat the generated crew as a strong starting point rather than a finished product, and recommend reviewing the task descriptions for places to be more specific. Everything built visually can also be exported as a Python project, and the same interface is available through APIs.
Why prompt detail matters
This is the clearest practical lesson in the session. A few sentences will get a crew started. The more detail you provide, the closer you get to something production-ready. Shane describes building a product marketing crew for competitive research and getting specific down to instructions like search Reddit for sentiment and find patterns of features known to have issues. At that level of specificity, the crew produced an analysis close to what would take him three days by hand.
Matching models to agents
In the demo, every agent starts on DeepSeek V3.1 Terminus. Justin then switches one agent to Llama 4 Scout, a smaller and more cost-effective model, and replays that single task to see the effect without rerunning the whole crew.
Shane expects more mixing and matching of models based on agent needs. An agent evaluating products across multiple dimensions may want stronger reasoning; an agent that sends a notification email does not need a state-of-the-art reasoning model. His rule of thumb: run the task with an expensive model and again with a smaller one, and if the results are around 95% similar, the smaller model is good enough. Some tasks genuinely need to be right 100% of the time, and those justify the larger model.
For choosing between models, the session recommends Artificial Analysis benchmarks for instruction following, coding, scientific reasoning and agentic tool use, then benchmarks specific to your own domain. Two other factors: whether the task needs a reasoning model, and per-model pricing, which varies considerably and adds up across repeated runs.
Why inference speed and cost matter for agents
A traditional assistant is a prompt template, an LLM and some retrieval: one request, one response. An agentic workflow is different. As the traces in the demo show, a single instruction can set off many LLM and tool calls in the background. Agents are chatty by definition, because they are reasoning, planning, exploring and verifying.
That changes the economics. Token volume per query has risen by roughly 10x to 100x as applications moved from chatbots to agentic systems, and the number of models involved has gone from one to five or ten. Shane compares the surprise to the early days of usage-based cloud pricing. The longer tokens take to generate, the more they cost.
Justin’s introduction covers why SambaNova is built for this. The RDU has two key innovations: a dataflow architecture that is more efficient than GPUs for inference and delivers performance at very low power consumption, and a three-tier memory design that supports both large models and multiple models on a single node. Citing an Artificial Analysis benchmark, he puts SambaCloud at roughly 10x faster than GPU-based inference.
He demonstrates it by asking for a Space Invaders game in HTML, giving ChatGPT a head start and still finishing first, at over 100 tokens per second with time-to-first-token under a second. Pasting the generated code into CodePen and running the game shows the output was accurate as well as fast.
Tools, triggers and guardrails
The session covers three categories of tools in CrewAI: open-source tools you use directly or with an API key, enterprise tools that require authentication for systems like HubSpot or SAP, and MCP server support, which was described as coming soon and will let you register servers and drag them into agents. Anything not available out of the box can be generated: Shane describes asking for a tool to create pages in Notion or retrieve records from HubSpot, and getting working code back in minutes.
Two other elements are worth planning for. Triggers make agents autonomous, letting them act on an event rather than waiting to be prompted. Guardrails keep agents from stepping out of line, whether that means inappropriate language or sharing sensitive information, alongside human-in-the-loop feedback for cases where a person has to approve the work. Use cases raised in the Q&A included competitive research, lead quality assessment, email triage, and appointment scheduling built from event-driven triggers plus calendar integrations. SambaNova also runs an early access program for startups, which includes credits.
FAQs
An agentic workflow uses multiple specialized AI agents that collaborate on a repeatable business process, rather than a single general-purpose assistant answering questions. The agents typically mirror real human roles on a team, are oriented toward tasks that involve many tools and steps, and operate with a high degree of autonomy, responding to changes and taking action without being prompted each time.
A few sentences is enough to generate a working crew. Beyond that, detail is the main lever on quality: the more specific the instructions, the closer the result gets to a production system. In the session, specificity down to which sites to search and what patterns to look for produced an analysis comparable to three days of manual work.
No. Models can be assigned per agent and individual tasks can be replayed to compare results. Agents doing multi-dimensional reasoning may need a stronger model, while an agent sending a notification email does not. A practical test is to run a task with an expensive model and again with a smaller one, and if the results are around 95% similar, use the smaller model.
A chatbot is one request and one response. An agentic workflow can trigger many LLM and tool calls from a single instruction, because agents reason, plan, explore and verify. Token volume per query has risen roughly 10x to 100x as applications became agentic, so slower generation means both longer waits and higher cost.