Skip to main content

Agenta Is Now a Workspace for Building Agents

Agenta started as a platform for managing and evaluating prompts. With this release, it becomes a workspace where you build agents: assistants that plan their own steps, call tools, and get work done, rather than answer a single prompt.

What an agent is made of​

An agent starts empty. You build it for a job by chatting with it, and the agent can write its own configuration as you talk. That configuration has five parts.

Instructions describe the job: what the agent does, who it serves, and how it should report back.

Tools are actions the agent can take in other systems, like creating a GitHub issue or sending a Slack message. Tools come from Agenta's built-in integrations, which cover mail, chat, calendars, documents, issue trackers, and CRMs, or from MCP servers you connect.

Skills are short handbooks for tasks the agent does sometimes. The agent loads a skill only when the task calls for it, so it isn't carrying instructions for every possible task on every turn.

Permissions decide which actions the agent takes on its own and which wait for your approval. You set a default for the agent, such as "read on its own, ask before writing," and you can override that default for individual tools.

Files live in the agent's folder: the drafts it writes, the notes it keeps, and the reference material you give it. You and the agent share this folder, so you can write documents or maintain a wiki together.

See Agents for the full picture.

Two ways to use an agent​

You can work with an agent in chat. You give it a task, it does the work, and you review the result together.

You can also run an agent as an automation, with nobody in the chat. An automation starts on a schedule, such as every weekday morning, or when something happens in an app you connected, such as a new row in a sheet or a new issue. The agent does the work and leaves the result somewhere you'll see it.

Choose your harness​

A harness is the program that drives the model: it sends the conversation to the model, runs the tool calls the model asks for, and feeds the results back. Agenta ships two harnesses, Claude Code and Pi, and you can switch harness or model without rebuilding the agent.

If you self-host Agenta, an agent can also run against a Claude or ChatGPT subscription you already pay for, instead of moving every task to metered API billing.

Versions and traces​

Every change to an agent's configuration is recorded as a version. You can read back what changed, when, and why, and compare versions to understand why an agent behaves differently than it did before.

Agenta also traces every run. You can inspect every model and tool call, and track requests, token usage, and estimated cost for each agent.

Open source​

Agenta is open source under the MIT license. You can inspect the code, self-host it, and run it in commercial projects.

Getting started​