Agents
A saved version of Asteria set up for one job, so you brief it once instead of every morning.
An agent is Asteria with a standing brief: its own instructions, its own documents, its own tools, and optionally its own schedule.
The problem it solves
You notice you are typing the same preamble every time.
You're reviewing copy for our brand. We're a B2B fintech. Tone is confident but never salesy. Never say "revolutionary" or "seamless". Check it against the style guide, flag anything off, and suggest a fix.
Once a week that is fine. Every day it is friction, and worse, you will do it slightly differently each time and get slightly different answers.
An agent is that paragraph, saved. You write it once, and from then on you just talk to Brand Voice Check.
When an agent is the right answer
Yes, build one when:
- You repeat the same setup, and the setup is more than a sentence
- Colleagues need the same job done the same way, and consistency matters
- The job always draws on the same documents
- It should run on a schedule with nobody watching
No, do not bother when:
- It is a one-off. Just ask in a chat.
- The instruction is short. "Reply in French" is not worth an agent.
- What actually changes every time is the documents, not the instructions. That is a collection.
Agents versus skills
These are the two things people most often confuse, and picking the wrong one is the single commonest reason a good idea ends up awkward to use.
An agent is who you are talking to. It takes over the conversation. You address it deliberately with @, and everything you say next goes to it until you exit.
A skill is how to do one thing. It does not take over anything. You drop it into whatever conversation you are already in, with /, and carry on.
| Agent | Skill | |
|---|---|---|
| Answers the question | Who is doing this work? | How is this step done? |
| How you invoke it | @ in the message box | / in the message box |
| Effect on the conversation | Takes it over until you exit | Applies and gets out of the way |
| Can carry collections, tools, a schedule | Yes | No |
| Usable inside any conversation | No, you have to be talking to it | Yes, including inside an agent |
The test that decides it
Read your instruction out loud. Is it a persona or a procedure?
"You are our brand voice checker. We are a B2B fintech, we write plainly, never say revolutionary." That is a persona, and it wants an agent.
"To format a case study: headline, then the client's problem in one sentence, then what we did, then the measured result, then a pull quote." That is a procedure, and it wants a skill.
Why it matters in practice
A procedure trapped inside an agent is only available when you are talking to that agent. The formatting rules you want to apply to a case study are just as useful in a chat about email, or inside a different agent, and as a skill they are available in all of those.
The reverse mistake is less costly but still annoying: a persona built as a skill has to be re-invoked on every single message, and it cannot carry collections, tools or a schedule.
They compose. An agent can have skills attached, so the brand voice agent can hold the case study formatting skill and load it only when a case study comes up. That is usually the right end state: personas as agents, procedures as skills, and the agent carrying the skills its job involves.
Talking to one
Type @ in the message box and pick the agent from the list. That is the only way to start a conversation with one.
A banner then reads "Talking to name" and stays until you click Exit agent. While it is up, that agent's instructions, documents and tools are in force and plain Asteria is not the one answering.
Clicking an agent's name in the sidebar does not start a chat with it. It opens that agent's settings, so you can change it.
This catches everybody once. To talk to it, go to the message box and type @.
The star beside each sidebar agent pins it to the top of the list. Underneath an agent with a schedule you will find its Trigger output thread, and clicking that does open a conversation, read-only, where its unattended runs post their results.
What an agent can carry
Everything from the builder, briefly:
| Instructions | The standing brief. The part that matters most. |
| Tools | What it may do: search the web, run code, read your mail, build documents. |
| Knowledge | Collections it always searches, plus files pinned into its context. |
| Skills | Procedures it can load when relevant. |
| Sub-agents | Other agents it can hand work to. |
| Triggers | A schedule, so it runs without you. |
| Suggested prompts | Starter questions, so a colleague opening it knows what to ask. |
Where they live
Agents in the sidebar. The page has your agents, agents shared with you, and a Browse tab for anything published to your organisation's marketplace.
Each card shows the agent's name, whether it is Private or Shared, how many tools it has, and a green dot if it is active. From the card: Edit, Duplicate, share, delete.
Next
- Building an agent walks through every part of the builder.
- Tutorial: your first agent builds a real one end to end in about twenty minutes.
- Sharing and the marketplace covers giving it to your team.