Shawn Swyx Wang argues that the agent opportunity will extend across the next decade, supported by large infrastructure investment and increasingly capable models.
2
Agent engineering now includes memory, planning, authority, control flow, tools, protocols, and the surrounding product interface, while memory and orchestration remain open areas.
3
Lélio Renard Lavaud says enterprise AI depends on organizing siloed data, adding context and observability, training people, and tailoring models and workflows to each business.
Summary
The opening session contains two talks about where AI engineering is heading. Shawn Swyx Wang gives a September 2025 update on agents. He defines an agent as more than an LLM in a tool loop, adding intent, memory, planning, authority, and control flow. He describes models that think with tools, run for longer periods, and receive more post-training compute. He also discusses MCP, A2A, ACP, agent APIs, coding agents, evals, context engineering, and faster inference. Lélio Renard Lavaud explains how Mistral approaches enterprise adoption. Companies have data spread across services, weak metadata, inconsistent access controls, and limited internal AI knowledge. Mistral works across models, APIs, products, and customer-specific solutions. Lavaud argues that useful systems need a context engine, sources that support trust, workflow observation, model improvement, and training. Open models provide a base, while continuous pre-training, post-training, deployment, and domain expertise create an enterprise business.
The agent opportunity extends far beyond a single year
Shawn Swyx Wang says the phrase "2025 is the year of agents" is too narrow. He points to an infrastructure buildout backed by what he describes as hundreds of billions of dollars, along with rapidly growing usage of consumer AI systems. Citing Andre Karpathy, he says the next ten years will involve building many agents. That gives engineers, startups, and businesses a rough direction for their work. His point is about duration and investment: the agent market is still early, even though products and model launches are changing quickly.
An agent needs memory, planning, authority, and control
Wang presents a broader view of agent engineering than an LLM calling tools in a loop. His list includes intent, memory, planning, authority, control flow, and tools. He argues that discussions should spend more time on memory, planning, and trust because these determine whether people can live with the resulting systems. The tool loop is only the visible part. A useful agent also needs a goal, a way to retain and retrieve information, a plan for acting, and boundaries on what it is allowed to do.
Agentic models increasingly think through tool calls
Wang describes agentic models as models that are "thinking with tools." Their reasoning can include tool calls as part of the process rather than treating tools as a separate wrapper around a finished answer. He also points to increasing autonomy, with systems moving from minutes toward hours and days of operation. Model providers are putting more compute into post-training so models can learn these behaviors. For application engineers, this changes what the surrounding software must manage, including long-running tasks, tool results, and model behavior over time.
Protocols and agent APIs are becoming part of the application layer
Wang says AI Engineer made a bet on MCP and asks what will be built on top of it next. He mentions Google's A2A and Zed's ACP, which he describes as an interoperability layer for terminal agents such as Claude Code and OpenCode. Developers are increasingly swapping whole agents and user interfaces instead of changing only a model name. He also describes an emerging standard library in agent APIs: code execution sandboxes, web search, document libraries, image generation, and MCP connections.
Agent platforms still lack good memory and orchestration
Wang updates Andrej Karpathy's language-model operating-system idea. Search, code execution, document libraries, multimodal input and output, and MCP are becoming established parts of the stack. He says good memory and orchestration remain unresolved, with several candidates but no clear winner. This leaves room for engineering work around how agents retain state, coordinate actions, select tools, and handle complex tasks. He presents these missing pieces as infrastructure problems rather than features that model providers have already settled.
Enterprise AI starts with data that has context and structure
Lélio Renard Lavaud says enterprise data has accumulated over decades across many providers, usually without metadata or an AI governance policy. Separate systems also have different access-control models. Retrieval alone does not solve this. His example is a query about a company's sock sales, where retrieval could return annual reports from the wrong years if the system does not understand the surrounding context. Mistral's proposed answer is a context engine that connects entities across data silos, maps them to an ontology, and creates an AI-first organization of the data.
Trust requires evidence, classification, and workflow observation
Lavaud says model outputs are inherently nondeterministic, so enterprises need tools that help people check and understand them. Sources can support verification, while classifying whether an output makes sense can detect divergence. Safety depends on the use case, since moderating Reddit comments differs from powering a children's chatbot. He also argues that companies need to observe how workers actually perform tasks, including the queries and workflows they use. That observation can reveal where AI can automate work or improve business processes.
Open models lead to enterprise work after pre-training
Lavaud says Mistral still expects pre-trained models to become mostly open and increasingly commoditized. The enterprise opportunity comes later, through continuous pre-training for languages or domains, post-training, distillation, deployment, and help with real customer data. Customers often lack the compute, engineering knowledge, or evaluation practice to do this alone. Mistral therefore combines open models with applied and solutions engineering. Lavaud describes the process as a loop: build, deploy, observe, improve, and repeat.
"Building a strong context engine that is able to build heuristics and that is able to actually understand the link between entities that might be residing in different data silos among different providers to actually make sense of this data is spread out everywhere."1:09:41
Who should watch
You are building an agent and need a current view of model capabilities, protocols, coding agents, context engineering, or inference speed.
Your enterprise AI project has data spread across SaaS systems and you need a concrete account of the governance, retrieval, observability, and training problems involved.
You are deciding whether to use open models directly or invest in domain adaptation, deployment help, and customer-specific engineering.
