The most useful shift in AI tooling this year isn't that any single tool got dramatically better — it's that the stack settled into distinct layers, each with its own job. Once you can see those layers, choosing tools gets easier, because you're picking the right thing for each layer instead of hunting for one product that does everything. If you want the broad lay of the land before the details, G2's emerging AI software overview is a decent snapshot of what teams are actually adopting; what follows is how the pieces fit together in practice. Coding assistants: part editor, part agent The first layer is the one most people mean when they say "AI tools," and it's changed the most. These have stopped being autocomplete and become agentic — reading repo context, planning multi-step changes, and running commands on your behalf. GitHub Copilot remains the low-friction default, the least disruptive way to add AI to an existing setup. Cursor is the AI-first editor of choice for serious full-stack work, though it rewards developers who review its output carefully rather than merging on faith. If you'd rather stay in the terminal, Claude Code and Aider lean agentic and shine when the task is less "finish this line" and more "plan this change and execute it across several files." The fork that actually matters when choosing is IDE-native versus terminal-agentic — and when a tool advertises itself as free, it's worth checking whether that means genuinely free or free-tool-but-you-pay-for-tokens, because those are very different for a daily driver. Application frameworks: building AI into your product A separate job entirely. This layer helps your application talk to models — handling prompts, tool calls, retrieval, and orchestration. LangChain and LangGraph cover the broad ecosystem, with LangGraph especially suited to stateful, multi-step agent flows. LlamaIndex is retrieval-first, a natural fit when your app is mostly answering questions over your own data. CrewAI and similar multi-agent frameworks handle setups where different roles hand work off to each other. The principle worth holding onto here is to keep this layer distinct from your editor and your model, so that when something better comes along you can swap a single piece instead of rewriting everything. Model gateways and routers You rarely want to hard-code a single provider. Gateways and routers let you work across the major hosted models and open-weight ones behind one interface, with fallback and cost tracking built in. OpenRouter gives you a single API across many models, which makes comparing and failing over straightforward. LiteLLM is an open-source proxy offering one OpenAI-compatible interface plus logging and budgets. The reason this layer matters: once you're building a product rather than just asking an editor for help, model choice becomes a cost and reliability decision as much as a quality one, and a gateway keeps that legible. Evaluation and observability This is the layer that graduated from nice-to-have this year. It lets you score outputs, catch regressions, and see what your application actually does in production rather than trusting that it looked fine in a demo. Tools like Braintrust and Maxim handle eval sets, CI runs, and live tracing. The mindset that pays off is treating every model output as untrusted until an eval says otherwise, and wiring those evals into CI the same way you'd wire unit tests. Retrieval and vector data The most mature and least churny layer. If you already run Postgres, pgvector is the place to start. Dedicated vector databases like Qdrant, Weaviate, and Milvus earn their place once scale or hybrid search justifies the extra infrastructure — but most projects reach production on pgvector, so there's little reason to add a database before Postgres genuinely can't keep up. How to vet a new tool The hype cycle is loud, so it helps to run everything through the same few filters: whether it fits your existing workflow rather than demanding you reorganize how you ship, how its output holds up on your own code rather than the demo repo, whether its cost stays legible when a hard task takes several attempts, and how it handles data and governance. For the quick "actually good or just well-marketed?" check, skimming verified developer reviews first saves a lot of time — G2's rundown of emerging AI software is one way to separate the tools with staying power from this month's launch noise. The takeaway is less about any one product and more about the shape of the stack: an editor or agent for writing code, a framework for the app logic, a gateway to keep model choice flexible, evals to measure quality rather than guess at it, and a retrieval layer for your data — five swappable pieces rather than one monolith you're locked into. What does yours look like? The gateway-and-eval combination is the part still shaking out, and I'd be curious what people have settled on.