Agents
Some work starts the same way every time. You gather the same background, follow the same process, and make the same judgment calls. In Agenta, you can turn that recurring job into an agent.
An agent is an assistant you set up for a specific kind of work. It begins each new task with the context, resources, and boundaries you have chosen, so you do not need to explain the job again from the beginning.
A repeated job becomes an agent
Imagine that your team launches a new feature every month. Each launch needs a brief that explains the audience, the problem, the message, and the channels you will use.
You could prepare every brief in a general chat and paste in the same company background each time. Or you could create a marketing launch agent that already knows how your team approaches a launch.
The useful boundary is the job, not a single request. One launch brief is a task. Preparing launch briefs for your team is a job that can belong to an agent.
Meet the marketing launch agent
When you ask the agent to prepare a brief for Team Workspaces, it does more than write from a blank page.
- Its instructions tell it who you sell to, how you position the product, and what a useful brief should achieve.
- A skill gives it your team's step-by-step process for preparing launch briefs.
- Its files hold product notes, brand guidelines, and examples of launches your team approved.
- Its tools let it check campaign data or create a task in another system.
- Its permissions let it research and draft on its own, while asking before it publishes anything.
These parts give the agent a stable way of working. Your request still supplies what is new: the feature, the deadline, and anything unusual about this launch.
Behind the scenes, a harness keeps the work moving and a model decides what to say or do next. You can choose both, but you do not need to understand their mechanics to start building an agent. See Harnesses and models when you are ready to compare them.
Create an agent through conversation or by hand
You can start by describing the job in the playground. Tell the agent what you are trying to accomplish, share examples, and correct its assumptions. As the job becomes clearer, the agent can help shape its own instructions and suggest the resources it needs.
You can also configure each part directly. Write the instructions yourself, choose the skills and tools, add files, and set the permissions. Both approaches build the same agent, and you can move between them as you work.
Build your first agent walks through the conversational path.
Feedback improves the next launch
Suppose the first brief speaks to small teams, but this product is for enterprise buyers. You explain the difference and add the target audience to the agent's instructions. If enterprise launches also need a security-review section, you add that step to the launch-brief skill.
When you commit those changes, Agenta records a new version of the agent. The next launch starts with the improved setup. You can compare versions to understand what changed or return to an earlier one if a change does not work as expected.
A session keeps one conversation together. A version records the agent's setup at a point in time. A new session does not carry over the old conversation, but it does use the version you selected. See Sessions and versions for the full distinction.
Over time, this feedback loop turns the agent from a broad starting point into an assistant that reflects how your team actually works.