Camila shares an AI agent roadmap

AI agents are having a curriculum moment. Not a new model launch, not a benchmark war — more like the community deciding what actually matters if you want to build useful agents instead of flashy demos. That’s the news here. Camila and a few other builders started passing around a practical roadmap on X, and the interesting part is what made the cut: not “learn prompt tricks,” but learn tools, memory, retrieval, orchestration, and production patterns. (anthropic.com) ### What was actually shared? The posts point people toward a cluster of resources that now function like a de facto agent syllabus: Anthropic’s “Building Effective Agents,” open GitHub repos like Nir Diamant’s GenAI_Agents, and community roadmaps that sequence the stack from LLM basics through RAG, multi-agent systems, and deployment. Basically, it’s a crowdsourced answe(anthropic.com)nt to ship agents that do work, not just talk about work? (anthropic.com) ### Why does this roadmap feel different? Because it treats agents as systems, not prompts. That sounds obvious, but a lot of beginner AI material still acts like the hard part is wording the instruction perfectly. Turns out the harder part is everything around the model — deciding when to call tools, how to store and retrieve context, how to break tasks into steps, and h(anthropic.com)ard into “simple, composable patterns” over giant abstract frameworks, which is a pretty strong tell about where the field has landed. (anthropic.com) ### What are the load-bearing skills? The roadmap themes are remarkably consistent across sources. First comes model and embedding literacy — enough to understand what the model is good at and what retrieval is fixing. Then tool calling and ReAct-style loops — the basic pattern where a model reasons, takes an action, observes the result, and keeps going. After that comes (anthropic.com)utorial repos, the progression is almost always from simple conversational agents to LangGraph or MCP tutorials, then to memory-enhanced and multi-agent systems. (github.com) ### Why is memory such a big deal? Because most useful agent tasks are not one-shot. A real agent needs to remember user preferences, prior steps, tool outputs, and sometimes long-running state across sessions. Without memory, you don’t really have an agent — you have a clever stateless assistant that keeps forgetting the plot. That’s why these roadmaps keep ele(github.com)production-focused guides too, where vector memory and stateful workflows sit right next to APIs, security, and observability. (github.com) ### Where does RAG fit in? RAG is the bridge between a smart model and the messy world of company data, docs, and current information. It gives the agent something better than vibes. But the catch is that RAG alone is not the product. The roadmap framing is useful here — retrieval is one subsystem inside a bigger loop that includes planning, tool use, memory, and evaluation. That(github.com)vector database.” (github.com) ### Why do multi-agent systems show up later? Because they’re usually not step one. A lot of community material now treats multi-agent setups as something you earn after you’ve mastered single-agent orchestration. That matches the more grounded advice from official guides — start simple, compose only when needed, and don’t add extra agents just b(github.com)urfaces get too large, but they also add latency, cost, and failure modes. (anthropic.com) ### So what changed for builders? The center of gravity moved. A year or two ago, “AI app” often meant prompt engineering plus a wrapper. Now the community’s most useful roadmaps are teaching agent engineering as workflow design: state, tools, retrieval, memory, evals, and deployment. That’s a sign of maturity. People are less interested in making the model sound smart and more interested in making the system behave reliably. (anthropic.com) ### Bottom line? The real takeaway from Camila’s roadmap moment is simple: if you want to build agents, study orchestration before aesthetics. Prompting still matters — but it’s no longer the whole game. The durable skill is learning how to make models act inside a controlled system, with memory, tools, and enough structure that a team can trust the output. (anthropic.com)