What it is
Not just "ask and answer" but an agent: the AI gets a goal, breaks it into steps itself, calls tools (search, databases, APIs, a browser), checks the result and drives the task to completion. The line is simple: a chat replies with text — an agent performs an action (creates a ticket, sends an email, updates the CRM).
Where it came from
The turning point was late 2024 → 2025: reliable tool use arrived, along with the MCP protocol (a single port between agents and tools) and Computer Use (the agent clicks with a mouse). Anthropic and OpenAI published guides on building effective agents, and in 2025 co-founded the Agentic AI Foundation under the Linux Foundation. By the end of 2025 there were 10,000+ MCP servers published.
Why it took off
The money went into applications, not infrastructure.
Per Menlo Ventures, businesses spent $37B on gen-AI in 2025 (3.2x year over year), and 51% of that went to the application layer. The reason: an agent addresses a concrete operational pain — the routine that used to require a person — rather than inspiration. ROI is counted in hours saved.
How to use it today
- Take one repeating chain (triaging inbound, research, a report) — not "the whole business at once".
- Describe the goal and the available tools; start semi-autonomous (the agent prepares, a human approves).
- Starter stack: a ready-made platform (n8n/Make plus an LLM node) or a framework (Claude Agent SDK).
- Measure deflection and time: what the agent closes by itself, and where it calls a human.
What to watch out for
Menlo's data says only 16% of enterprise deployments are "real agents"; the rest are fixed scripts around a single model call. Don't confuse automation with agency. Grant narrow permissions, log every action, and keep a kill switch for anything irreversible.