- Home
- Skills
- Workflow Automation
- Backend Building Swarm: Parallel Agents, One Green Main
More screenshots
Works with the AI Agents you already use
Backend Building Swarm: Parallel Agents, One Green Main
Build the backend with parallel agents in isolated git worktrees: worktree discipline, per-agent build contracts, and a merge-back protocol that keeps main green.
$5
Backend Building Swarm: Parallel Agents, One Green Main
Example session with this skill installed
Build a project-tracker backend with three parallel agents: A does auth + users, B does billing + webhooks, C does domain routers (projects, tasks, tags). Main is green at the base commit. Coordinate the swarm and merge everything back green.
- Read your context and instructions
- Compiled the backend building swarm
SWARM COMPLETE — 3 slices merged, full gate green on main, cross-slice integration test passing.
Contracts written before spawning: Agent A owns the users table and all auth procedures; it provides auth.signUp, auth.signIn, and the session context every other slice consumes. Agent B owns billing tables and the webhook route; it consumes the users table read-only through A's interface. Agent C owns projects, tasks, and tags; it consumes the session context from A. Schema: agent A owns all schema changes (centralized — the simplest choice for three slices).
Build loop: each agent works in its worktree, rebasing onto main daily. On day two, agent C needs a plan field on the user to gate project creation — a schema change. C does not edit the schema; it files a contract amendment with the orchestrator, who routes it to A. A adds the field, C consumes it through the interface. No cross-worktree import, no conflict.
Merge-back (dependency order): A finishes first — diff reviewed, merged, full gate green on main. B next — merged after A (billing references users), gate green. C last — merged, gate green, and the orchestrator's integration test exercises the cross-slice path (sign up → create project → check plan gating) on the combined tree. It passes.
A failure, handled: on B's merge, the gate fails — the webhook route imports a helper that A renamed in its own slice. The orchestrator traces the failure to B's slice (A is green on main), sends B the failing output, and B fixes the import against the merged main. One rebase, one re-test, gate green. The failure was caught because the gate runs on the worktree rebased onto main before merging, not after.
Connects securely to your tools. The creator never sees your data.
About this skill
One agent building a backend is slow; five agents editing the same tree is chaos. Worktrees solve it: each builder gets an isolated checkout of the shared repo, builds its slice, and merges back through a contract. The orchestrator's job is contracts and merges, not code.
Each worktree is a full git checkout on its own branch, created from the same base commit, with dependencies installed inside it — never shared across worktrees. Before spawning, every agent gets a written contract: scope (exactly which routers, tables, and files it owns — overlaps are forbidden), interfaces (the tRPC procedure names, input shapes, and tables it may read or write), done criteria (builds clean, its tests pass, a short note on what was built and what it assumes about other slices), and branch name plus base commit. The schema gets special treatment: one agent owns all schema changes, or the schema is split by table prefix with zero overlap.
Inside its worktree, each agent works like a solo builder, with two swarm rules: rebase onto main daily, and no cross-worktree imports (a need from another slice is a contract amendment through the orchestrator, not a relative import). The orchestrator merges — agents never merge each other's work — in dependency order (schema/migrations first), and runs the full gate on main after every merge: typecheck, full test suite, build. If the gate fails, the culprit slice goes back with the failing output, not a vague "fix it." A thin integration test on main exercises cross-slice paths at every merge, catching the silent coupling that per-slice tests miss.
- A swarm without an orchestrator is chaos with more participants — someone must own contracts and merges for the whole run.
- It is overkill for a backend one agent can finish in a day; the coordination cost exceeds the parallelism gain.
- Heavy shared-state coupling between slices defeats the split — the slices must be separable with clean interfaces.
- A worktree should live hours, not weeks; a slice that takes days while main moves on produces a merge nobody can review.
- Interfaces are zod schemas in a shared file, not prose — prose contracts drift.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 3 days ago
Permissions
No special permissions declared or detected