Skip to main content

The Pattern

Two autonomous agent workflows (FORGE and TOMMY) were validated end-to-end through a BFF webhook trigger pipeline. FORGE selects the next unblocked ticket from a GitHub Project board and generates an implementation prompt. TOMMY atomises a repository into parallelisable task-lets. Both accept parameterised input (org, repo, branch, filters) and return structured responses. The key architectural insight is that n8n webhooks serve as an Agent-to-Agent (A2A) protocol boundary. The BFF doesn’t need to understand what the agents do internally. It discovers the webhook path from the workflow definition, forwards the payload, and returns the response. This makes adding new agent workflows a zero-code operation on the BFF side.

Cross-Org Scalability

Both workflows proved cross-org capable in testing. FORGE was tested against traceo-ai/traceo-mcp-server (primary target) and so1-io/so1-content (cross-org probe). TOMMY was tested against both orgs and returned different tasklet counts based on repository structure. The only constraint is whether the GitHub token in n8n has access to the target org. No code changes are needed to support new orgs.

Operational Model

The validated pipeline: Rover UI form -> BFF proxy (BetterAuth session + trusted headers) -> BFF trigger route (webhook auto-discovery) -> n8n webhook (120s timeout) -> structured response displayed in UI. Schema-driven forms mean FORGE and TOMMY get purpose-built input fields (dropdowns, toggles, number inputs), while unknown webhook workflows fall back to a raw JSON editor. This gives non-technical operators access to agent workflows without touching the n8n console.

What This Unlocks

With FORGE and TOMMY triggerable from the Rover console, the next step is chaining them: FORGE selects a block, TOMMY atomises it, and a third workflow (or direct SSH execution) runs the task-lets. This creates a fully automated sprint execution loop governed by the PROJECTS.md execution checklist, where human intervention is only needed for review and approval.