Agents in Production: How AI Started Doing Instead of Saying

For most of AI's commercial history, the dominant interaction pattern was query-response: a human asked a question, an AI answered it, and the human acted on the answer. In 2025, that pattern began shifting at scale. AI agents (systems that break a goal into steps, use tools to execute those steps, observe results, adjust plans, and complete multi-step tasks without per-step human intervention) moved from research demonstrations and narrow pilots to meaningful production deployments across industries. The shift was not sudden; it was the convergence of several enabling technologies crossing thresholds simultaneously.

The ingredients had accumulated over three years. Function calling gave LLMs the ability to invoke external tools deterministically. Vector databases and RAG gave them access to organisational knowledge. MCP (November 2024) standardised tool integration across frameworks. Long context windows (Gemini 1.5 Pro at one million tokens, Claude 3.5 at 200k) let agents hold large codebases or document collections in working memory. Reasoning models (o1, o3, R1) improved multi-step planning reliability. Eval frameworks (SWE-Bench, GAIA) provided measurable benchmarks for real task completion. Each development was necessary but not sufficient. Together, they crossed a threshold where agentic workflows became viable outside controlled demonstrations.

The agent frameworks that matured through 2024–2025 each attacked the core problems of agentic systems from different angles. LangGraph built explicit state machines for agent workflows: making execution paths inspectable and debuggable. CrewAI specialised in multi-agent role assignment, allowing different agents to own different responsibilities within a shared task. AutoGen focused on conversational agent collaboration with human-in-the-loop checkpoints. Anthropic published agent design patterns emphasising minimal footprint and explicit confirmation before irreversible actions. No framework solved all the problems cleanly. The diversity reflected genuine uncertainty about which abstractions would hold under production load.

The gap between agent demos and agent production turned out to be almost entirely about failure handling. Demo agents succeeded on the happy path: well-specified goals, working tools, unambiguous intermediate states. Production agents encountered underspecified goals written by non-technical users, APIs that returned unexpected errors, tool outputs that contradicted each other, and edge cases that fell outside the scenarios the agent had been designed for. The production failure rate of first-generation agents was high enough that organisations that shipped successfully shared a common pattern: they defined the narrowest possible task scope, instrumented every execution step, measured failure modes systematically, and expanded scope only after the narrow case proved reliable in production.

The enterprises that deployed agents successfully in 2025 treated the agent like a junior employee: capable within a defined scope, requiring clear escalation paths when uncertain, and supervised rather than autonomous on consequential actions. "Automate invoice matching" worked. "Handle all accounts payable" did not: not because the technology was inadequate for the first task, but because the second task had not been scoped into components the agent could handle reliably. The recurring lesson was that the capability to complete a task in a demo and the reliability to complete it in production at scale were two different things, separated by engineering work that no benchmark measured and no model could substitute for.