AI agent

An AI agent differs from a chatbot in one structural way: a chatbot answers the message in front of it, while an agent works toward a goal across multiple steps. The common pattern is a reasoning loop — often called ReAct — where the model writes down a thought, picks a tool, calls it, reads the actual result back, and repeats until it reaches an outcome or a limit.

Why it matters

The same loop that makes agents useful makes them risky. An agent that can call tools can also call the wrong tool, loop indefinitely, or burn tokens on a dead end. The questions that matter when evaluating any agent product are mechanical: what bounds a run, who approves a write, and what record survives afterwards.

How agents work in Oneop

Oneop’s AI Workforce runs real ReAct agents against your CRM, tickets, tasks and knowledge base. Every reasoning step goes through one metered routing function, so each run charges the workspace’s AI credit balance and stops at the cap rather than billing overage.

Four hard ceilings bound every run: six reasoning iterations by default with an absolute cap of twelve, a 6,000-token budget, a forty-five-second wall clock, and three actions per run. Hitting any ceiling ends the run as escalated, with the exit reason recorded — the agent does not invent an answer to finish.

When an agent proposes a write, it does not execute it directly. The action routes through a governed engine that applies the agent’s autonomy level: log it, queue it for a human, or dispatch it. Actions that touch money or stock are held behind a separate switch that is off unless it has been turned on for your workspace, regardless of the autonomy setting.

Every run stores its full trace — each step’s thought, chosen tool, actual observation, token count and guardrail result — and is badged Live or Simulated. A simulated run makes no model call and writes zero credit usage, so a demo can never be mistaken for real work.

All agent reasoning runs on Google Gemini, Oneop’s sole model provider.

Related terms: autonomy levels, human-in-the-loop. See the product page: AI agents.