Recipes#
Each recipe builds one real thing from start to finish: the choices, the specification, the steps in the console and the API, what you should see, and how to check it. Start with the one closest to what you are doing.
-
A pull-request reviewer with an approval before merging
Reviews a pull request on GitHub, comments freely, and merges only after an admin approves.
Uses: the GitHub connection,
@approval, a guardrail, receipts. -
A support agent on an Eos model
Answers from your Notion knowledge base and drafts ticket replies, on a model served on your own GPUs: no prompt leaves your machines.
Uses: an Eos deployment, Notion read-only, an HTTP tool with writes needing approval.
-
A research agent with the web tool and a budget
Reads public pages and answers with sources, stops at $1 a run, and asks before passing its monthly budget.
Uses: the web tool, the provider's search, run and period budgets, model routing.
-
Exports yesterday's incidents, has an agent summarise them, asks, and publishes — every night, surviving a machine going away.
Uses: flows,
retry,when, a shared drive, an approval step, a scheduler of yours. -
Guardrails that forbid destructive tool calls
No agent of the workspace deletes, drops, pays or sends outside office hours, whatever its own policy says.
Uses: workspace guardrails, policy checks, organisation guardrails.
-
Evaluate and promote a new agent version
Build a suite from real runs, require it, give the new version 10 % of runs, then all of them — or roll back.
Uses: eval suites, graders, the gate, promotion, traffic, rollback.
-
Give an auditor a quarter's governance record, signed and checkable offline, mapped to EU AI Act and SOC 2 controls.
Uses: the auditor role, retention, evidence packs, receipts.
How the recipes are written#
- Every step shows each way to do it, in the order Console, CLI, API.
astracovers flows and verification (astra anemoi,astra evidence); agents, approvals, budgets, guardrails and evaluations are done in the console or with the API. -
API calls go through the console with a personal API token, to your workspace on a cluster. Each recipe assumes:
-
Machines are Linux machines of the workspace that can run agents (kernel 6.2 or newer, the current machine package). See Agents, runs and sandboxes.