Build a workflow from scratch

Use this guide when you want to create a repeatable workflow instead of running one ad hoc request from chat.

Before You Start

Start the dashboard at http://127.0.0.1:4317/.
Install and launch Aegis from Installation.
Prepare the target project with Prepare a project for Aegis.
Configure at least one model provider.
Configure any connector accounts the workflow will need.
Decide what should trigger the workflow: manual run, schedule, webhook, or another event.

Create The Workflow

Open the dashboard.
Go to the workflow designer.
Create a new workflow.
Give it a short name and a clear objective.
Add a trigger node. For a first workflow, use a manual trigger.
Add one or more work steps:
- agent step for reasoning, coding, research, QA, or planning;
- connector tool step for Gmail, Calendar, Slack, GitHub, HTTP, or another app;
- parser/filter/branch/merge/loop/retry step when input needs routing or repeatable control flow;
- storage or artifact step when a later node needs durable intermediate data;
- source, sandbox, quality preflight, report, or delivery step for report workflows;
- notification step when the workflow should report back.
Connect the nodes in the order work should run.
Configure each node from its settings panel.
Save the workflow.
Run a small test with safe input.

Configure Agent Steps

For each agent step, set:

role or profile;
provider and model;
prompt or objective;
allowed tools;
connector accounts and scopes;
expected output shape;
approval policy.

Keep the first version narrow. A workflow that does one thing reliably is easier to inspect, approve, and improve.

Add Approval Gates

Use approval gates for actions that write files, call external systems, send messages, publish content, modify accounts, or change workflow authority.

For repo writes, Aegis should create a patch bundle and verifier output before asking for approval. Review the diff, hashes, and policy class before approving.

Test The Workflow

Run the workflow with a low-risk input first.

Check:

the run reaches the expected status;
task sessions were assigned to the right roles;
artifacts were created in the expected paths;
verifier output is present;
approvals appear only where expected;
notifications are useful and not noisy.

Improve It

After the first run, adjust:

prompts that were too broad or vague;
connector scopes that were too permissive;
missing branches or retries;
output names and artifact paths;
approval policies;
dashboard notifications.