Everyone says they have an agent.
Almost no one means the same thing.
The word has gotten slippery. It now covers everything from a chatbot in the corner of your screen to a system that can knock multiple to-do’s off your list in an afternoon. This is a short guide to telling agents apart, and how to spot the real thing when buying software.
The word “agent” is used a lot. But what is an agent?
Strip away the marketing and the idea is simple. For decades, software waited for you. You clicked, it responded. You typed a number, it stored it.
An agent is meant to flip that around. Instead of waiting for the next click, it takes a goal and works toward it: deciding the steps, doing them, noticing when something’s off, and coming back to you only when it needs guidance or is done.
An agent is software you can hand a task to and reasonably trust to carry it out, flag what’s unclear, and tell you when it’s done.
That’s the promise. The term “agent” has become confusing because almost all products are using the word, and most stop well short of that definition – only handling a small number of tasks or unable to take action. It’s worth knowing the types of agent you’ll see in the wild, and which ones are worthy investments.
Most things called an “agent” are really one of three things.
You’ll probably recognize all three. Each offers some value, with real limitations. Once you see the pattern, you’ll know to ask for more.
All three are helpful. But notice the pattern: each only handles a discrete task and, at the end, the work comes back to you. Each makes one step easier, but can’t take the next step.
True agents can do all of this and more, carrying the work through to the end.
When you hear “agent,” ask: Can it do everything I can do in the software, or just what you demoed?
A true agent is general purpose. It can do anything a person on your team can do in software, with code instead of clicks. There’s a big gap between an agent that can do almost anything on a platform, and an agent that can only do the three things it was wired to do.
If it only works on a handful of tasks they demoed, it’s a feature, not an agent.
To understand the difference, it helps to see one. So here’s the Gravity agent doing a job, from start to finish.
An all-purpose agent that does the work, from start to finish.
The Gravity agent can do most anything a user can do in the platform. Here’s an example of a real job we handed to the agent and what it did. Scroll to watch it work through the problem, step by step.
You hand it a goal, the way you would a new hire.
First it breaks the goal into a task checklist.
It goes and finds what it needs: files, your data, web searches.
It reads last year’s report so this one matches.
It shows you exactly what it wants to bring in, for approval.
It drafts your disclosure, in your own voice.
You approve and it saves the draft, with a record of all steps taken.
Gravity Agentworking · live Ready This is only possible when software is “API-first.”
When a platform is built this way, the API serves as a door that lets the software use every tool in the platform, so agents can do a wide variety of work. When a platform isn’t built this way, the vendor is forced to bolt shortcuts onto an existing product and call each one an “agent.”
What makes all-purpose agents useful? Expertise embedded in “skills.”
Many agents are a thin layer over an off-the-shelf AI like ChatGPT or Claude, dressed up with a clever prompt. That’s fine for narrow tasks, but the underlying AI lacks the expertise to reliably do complex work. To be useful, agents require skills – reusable expertise that the agent applies automatically when relevant.
A skill is detailed written guidance that an agent reads before doing a specific kind of task, including both instructions and context.
At Gravity, for example, we have teams of full-time climate and energy experts who continuously create new skills for the agent. We have trained 85+ skills so far, which codify expertise to migrate data from Excel and other platforms, identify cost-saving energy projects, and draft reports for regulations like California’s climate disclosures, among other use cases.
Here’s how an all-purpose agent is trained:
A real agent isn’t a thin wrapper over an off-the-shelf AI. It embeds real expertise, consistently updated and validated by human subject matter experts.
“But is it safe?” Here’s what to look for.
Agents are a new way to work with software, and they naturally raise questions around control, auditability, and access. Enterprise-grade agents will have guardrails in place to address your concerns.
A mature agent should operate like a new team member with a login: it works inside the same access controls and security measures your team already trusts. It can only see and do what the person running it is allowed to see and do.
A Gravity agentSees only what its operator allowsEnterprise-grade agents also offer additional guardrails:
Approvals
Agents should always request human review and approval before taking action, for example editing data or drafting a report.
Data security
Agents should only access the data the person using them can access, or a subset of that data.
Auditability
Agents should always leave a trace, creating an audit trail so users can review what happened, when, and why.
Access controls
Agents should be subject to the same permissions as employees. They should never get a master key.
Nine questions worth asking when evaluating an agent
The answers reveal whether you’re evaluating a true agent, or just a feature.
not on slides.
Bring your data. Watch it work.
The fastest way to tell a real agent from a feature is to put your own data in front of it. Book a short call and we’ll do exactly that, live.