What is an AI agent?
An AI agent is software that takes an objective and autonomously executes the steps to complete it — across tools, over time, with memory. A chatbot responds to prompts; an agent works on goals.
The phrase 'AI agent' has become the most overloaded term in software. Every vendor claims to ship one. Most ship a chatbot. This post draws the line between the two and walks through what an AI agent actually is — what it does, what it remembers, and what it decides on its own.
The three properties of an AI agent
An AI agent, in the precise sense the term is used by practitioners, has three properties that a chatbot doesn't:
- Goal-directed execution. You give it an objective (e.g., 'source 10 qualified candidates for this role by Friday'), not a prompt. It decomposes the goal, picks tools, and runs the steps.
- Autonomy across steps. Between input and output, an agent may take 10, 100, or 1,000 steps — planning, tool-using, self-correcting — without asking for permission at each one.
- Persistent memory. An agent remembers the context of prior work. It picks up where it left off, applies learned preferences, and avoids re-asking what it already knows.
A chatbot has none of these. It takes a prompt, returns a response, and forgets the interaction. It is a turn-based conversation. An agent is a background process.
What an AI agent actually does
To make this concrete: an AI agent for a recruiter might receive a job description on Monday morning. Between Monday and Friday, without further prompting, it does the following:
- Parses the JD for must-haves, nice-to-haves, and red flags
- Sources candidates from LinkedIn, GitHub, and the firm's ATS
- Scores each candidate against the JD using the recruiter's specific criteria
- Drafts personalized outreach messages in the recruiter's voice
- Schedules intro calls when candidates reply
- Updates the ATS continuously with status and notes
- Surfaces the top 10 candidates each morning for the recruiter's judgment
That's an agent. The recruiter is never in the middle of the loop — they're at the end of it, making the decisions that require human judgment.
"The test: if you have to prompt it more than once to get output, it's a chatbot. If you set a goal and it produces finished work, it's an agent.
— Spawnlabs, on the definition
Why the distinction matters
Chatbots are net-neutral on knowledge work. They answer faster than Google; they don't change throughput. Agents are net-positive — they execute work that would otherwise require human hours. A recruiter with three agents isn't a faster recruiter; they're a recruiting team.
This is why enterprise AI budgets are shifting: away from chat-UI seats and toward agent licenses. The ROI math is different. A chat license saves minutes; an agent license replaces a function.
The architecture under the hood
Most working AI agents today are built on a few primitives:
- A reasoning loop — the agent plans, acts, observes, and iterates (commonly called ReAct or one of its variants)
- Tool use — structured APIs the agent can call (search, email, calendar, CRM, custom SQL)
- Memory — a persistent store (often a file like MEMORY.md or a vector DB) that survives across sessions
- Sandboxed execution — an isolated environment where the agent can read, write, execute code, and recover from failures
- Skills or playbooks — reusable process definitions that the agent applies to specific tasks
None of these are new individually. The leverage comes from stitching them into a reliable loop that runs in the background without a human at the keyboard.
The Spawnlabs take: agents represent people, not roles
The framing we think matters most: an AI agent should represent a specific person, not a generic role. A 'recruiting assistant' is a chatbot. Sarah's recruiting agent — encoded with how Sarah specifically screens, sources, and sells — is the thing that scales.
That's the bet Spawnlabs is making. Agents are owned, portable, and shareable by the expert. A firm that adopts Spawn doesn't get 'a sales agent.' It gets an agent for its best salesperson, and another for its best operator, and another for its best support lead — each encoded to a real human's way of working.
Next read: What's the difference between an AI agent, a chatbot, and an assistant? We break down all three.