A standing agent runs on a schedule, writes a report and raises suggestions. This page is about adding and configuring one.
Add an agent#
On the Agents page you can:
- Add one from the catalogue. The agents Eido Build ships with are ready to use and can be changed afterwards.
- Press New agent and choose between describing the agent to Odie, or building it by hand in the editor.
Building one with Odie#
Describe the agent in a sentence. Odie then:
- Asks questions until you both agree what the agent is for.
- Checks what integrations already exist, and reads the documentation of any API the agent needs.
- Proposes each piece in order: a secret for the credential, the integration, its actions, and then the agent with its grants.
Every step is a card that you press. If a credential is needed, Odie tells you where to create it, and you paste it into a card that sends it straight to Eido Build’s secrets. The value never enters the conversation.
What you set on an agent#
| Setting | What it does |
|---|---|
| Name, icon and colour | How the agent appears. |
| Instructions | What it is asked to do on each run. |
| Project | Which project it works on. Its brief includes that project’s description. |
| Schedule | By hand, hourly, daily, weekly, monthly, when a release is cut, or when a work item finishes. Times are in UTC. |
| Workbench and model | Where it runs and on which model. |
| Report | The name of the document it writes each run. |
Budgets#
| Cap | What it does |
|---|---|
| Per run | The run stops when it has spent this much. |
| Per day | A run that would take the day’s spending over this does not start, and is recorded as skipped. |
Both are in US dollars. See Costs and budgets.
Code access#
| Setting | The agent can |
|---|---|
| None | Not see the repository. It works from what Eido Build knows. |
| Read | Read the repository. It is given a read-only token, so a push is refused. |
| Write | Branch, commit and push. Off unless you turn it on for this one agent. |
Trust, one box at a time#
Everything here is off to start with.
May raise work items. The agent writes the work item itself instead of proposing one. It still lands in Triage for a person to approve. This suits an agent whose whole job is filing what it found.
May add customers. For agents that read where requests come from and keep finding names nobody has written down. Anybody an agent adds is a prospect.
Integration actions. Which actions the agent may perform in other systems: none, chosen ones, or everything including actions added later. This is the only way an agent reaches another system. See Integrations for agents.
Memory#
Turn memory on for an agent that comes back to the same job every week. It gets a folder that survives between runs, and reads and writes its own notes there. Turning memory off keeps the notes.
Skills#
A skill is a folder of instructions and reference files that an agent can use: a SKILL.md and whatever it refers to. Upload the zip you already have, or type one in.
If a skill cannot be delivered to where the agent runs, its text is added to the agent’s instructions instead, so the run still happens. The Skills screen says which skills were delivered.