Agents & tools
Autonomous Agent
An AI agent that operates with minimal human oversight — making and executing decisions independently across long time horizons.
01 ——
In plain English
An autonomous agent is an AI agent designed to run on its own, often for long periods or many steps, without a human reviewing each action. Where a copilot suggests and a human accepts, an autonomous agent decides and acts.
Spectrum of autonomy:
- Suggest-only (autocomplete, smart compose) — human approves every action
- Assist (copilots) — human accepts/rejects model proposals
- Supervise (most current agents) — agent acts, human reviews at checkpoints
- Autonomous (some research agents, trading bots) — agent runs end-to-end, human reviews outcome
Why pure autonomy is rare:
- Models still hallucinate or get stuck in loops
- Mistakes compound over long runs
- Liability and trust haven't caught up with capability
- Human-in-the-loop checkpoints catch most failures cheaply
Where you see it today: Long-running research agents (Deep Research), automated trading systems, security scanners, and self-hostable open-source agents (AutoGPT lineage). Most "autonomous" production agents still have at least one human gate.
02 ——
Related terms
AI Agent
Software that uses an LLM to plan and act — picking tools, taking actions, and adapting based on results to complete a user’s goal.
Agentic AI
AI systems designed to act, not just respond — they plan, use tools, and make decisions across multiple steps to complete a goal.
Agentic Workflow
A multi-step task where an AI agent autonomously decides the steps, uses tools as needed, and works toward a goal with minimal human steering.
Human in the Loop
An AI workflow that pauses for a human to review, approve, or correct the model’s output at key steps — instead of running fully autonomously.
Agent Orchestration
The coordination layer that decides which agent or tool runs next, manages state across steps, and handles failures in multi-step AI workflows.
Agent Harness
The scaffolding around an LLM — tools, memory, loops, and orchestration — that turns a model into an agent.
Back to glossaryLast reviewed May 2026