# AI Agent

> VI: https://5ac.vn/wiki/ai-agent/ · EN: https://5ac.vn/en/wiki/ai-agent/ · lang: en · spec_as_of 2026-10-19 (spec no formal spec) · cluster C2

Software that uses a large language model as its brain to plan, call tools and finish tasks toward a goal — proactive where a chat assistant is reactive.

---
## Definition

An AI agent is software that uses a large language model (LLM) as its brain to pursue a goal: it breaks work into steps, decides which data sources to read and which tools to call, checks the result and moves on without waiting for a human to issue each command. 5ac separates agent from assistant by proactivity. A 2025 assistant was reactive: you asked, it answered. A 2026 agent is proactive: it monitors the inbox, qualifies leads, drafts replies and flags the owner only when something genuinely new appears. In production an agent stops being a chat window and becomes an operating layer: one task may call several models, read several data sources, run tools and accumulate cost step by step.

## Mechanism

The scholarly frame is *A Survey on Large Language Model based Autonomous Agents* (arXiv:2308.11432, Lei Wang et al., 2023), which models LLM agents as systems with planning, memory and an action interface to their environment, not just text generators. The 5ac articles organize the moving parts into three layers:

| Layer | Contents | What to watch |
| --- | --- | --- |
| Action loop | plan, fetch data, call tools, verify, repair, re-run | one request can become dozens of model calls |
| Access | the data sources and tools an agent may touch | what it read, what it changed |
| Operations | token budgets, model routing, logs when many agents run in parallel | cost per finished task, not per call |

The way 5ac organizes agents inside a company is by role, not by model: a content agent reads briefs and brand docs, a CRM agent touches the pipeline, a review agent only checks output, and a human approves decisions that are hard to undo. A "sovereign agent" is this same entity judged by control instead of speed: where the data lives, who holds access, which actions need approval, whether the logs support an audit.

## Application context

For small businesses, agents enter through repetitive workflows close to revenue: handling 70–80% of Tier-1 support questions, qualifying leads, drafting first responses, producing internal reports — jobs a ten-person team cannot staff around the clock. Three operating costs ride along, per 5ac's orchestration piece: a token budget per workflow, role-based access, and a log for every action. Without them, early results arrive fast while the bill grows quietly.

## Sources fetched

- [AI Tools for Small Business 2026](/en/blog/ai-for-smb/) — reactive assistant vs proactive agent, Tier-1 figures — accessed 2026-10-19
- [Agentic AI orchestration: token cost control for SMBs](/en/blog/agentic-ai-orchestration-token-cost-control-smb/) — the agent as an operating layer with per-step cost — accessed 2026-10-19
- [Sovereign AI agents for Vietnamese SMBs](/en/blog/sovereign-ai-agents-vietnam-governance-compliance/) — control judged by data, access, approval and logs — accessed 2026-10-19
- <https://arxiv.org/abs/2308.11432> — the autonomous-agents survey (v7) — verified via arXiv API 2026-10-19

## Related

- [Agent governance](/en/wiki/agent-governance/)
- [What is a sub-agent?](/en/wiki/sub-agent/)
- [Agent memory](/en/wiki/agent-memory/)
- [Agent skills](/en/wiki/agent-skills/)
- [Model Context Protocol (MCP)](/en/wiki/mcp/)