Workflows
A workflow in eidan is a repeatable, staged process the agent runs — some steps are deterministic tools (no model), some are bounded LLM steps (grounded, short-context). eidan ships a growing library of them; this section explains each with a simple diagram so they can be reviewed and improved.
The shape of every workflow
eidan runs on two layers: a stable layer of deterministic tools/plugins, and an evolving layer of LLM-driven steps (workflows + skills) that call into it. A good workflow pushes as much as possible down to the stable layer and keeps the LLM steps short and grounded.
The workflows
| Workflow | What it does |
|---|---|
| Content | Plan a campaign on a board: concept → assets → copy → review → schedule, each column a gated step. |
| Sage (coding) | Delegate an engineering task; it branches, codes, opens a PR, and iterates to green. |
| Agents & triggers | A user-defined agent fires on a schedule/sensor/webhook, runs one turn, escalates when unsure. |
| Reasoning | Skills are classified per turn and loaded on demand; cognition adds inner-voice critique. |
| Rethinking | Resolve unknown terms just-in-time (rumsfeld); get a second opinion before committing. |
These are review drafts
The diagrams are meant to be read, argued with, and refined — if one doesn't match how a workflow actually behaves, that's a doc bug worth fixing. Edit the page on GitHub.