For the first years of the LLM era, AI was something you talked to. The defining shift of the current phase is AI you delegate to — agents that take a goal, make a plan, use tools, and carry multi-step work to completion.
What changed
Three ingredients matured together. Models got reliable enough at tool calling — emitting structured requests to run code, query APIs, or search. Research patterns like ReAct showed how to interleave reasoning with actions so a model can check its work. And frameworks such as LangGraph made these loops controllable enough for production, with state, retries, and human checkpoints.
Where agents actually work today
Coding is the breakout. Agentic coding assistants take a ticket, explore a codebase, write a change across files, run the tests, and open a pull request. The pattern generalizes: the work is digital, verifiable, and reviewed before it ships. Customer-support triage, research and summarization pipelines, and back-office data workflows share those traits and are following the same adoption curve.
The honest constraint
Reliability compounds badly. An agent that is 95% correct per step is only ~60% reliable across ten steps. Today’s successful deployments accept this and design around it: narrow scopes, verifiable outputs, and human approval gates at consequential moments. Full autonomy remains the exception, not the rule.
What to watch
Three signals for the next phase: standardized protocols connecting agents to tools and data; agents charged per outcome rather than per token; and enterprise platforms making “human-in-the-loop by default” the deployment norm. The direction is clear — the interesting question is how fast the reliability gap closes.