Solo Dev Team: Complete OpenClaw Agent Swarm Configuration Guide

Original Author: Elvis (@elvissun) Source: X/Twitter Original Link: https://x.com/elvissun/status/2025920521871716562

Core Architecture

The author no longer uses Codex or Claude Code directly, instead utilizing OpenClaw as an orchestration layer. His AI assistant, Zoe, is responsible for:

Generating Agent prompts
Selecting the appropriate model for each task
Monitoring progress
Notifying him on Telegram when PRs are ready to merge

Performance Metrics

94 commits/day — Most productive day: 3 client meetings held, editor opened zero times
7 PRs/30 mins — From idea to production at extreme speed
Almost all small to medium tasks are completed in one shot

Why do you need an Agent Orchestration Layer?

The context window is a zero-sum game:

Fill in code → No space left for business context
Fill in customer history → No space left for the codebase

This is the advantage of a two-layer system: each AI only loads what it needs.

8-Step Workflow

Step 1: Client Request → Zoe Evaluates

Automatically syncs meeting notes from Obsidian with zero context-switching cost. After a client call, discuss requirements directly with Zoe; she automatically retrieves relevant background information.

Step 2: Spawn Agent

Each Agent has an independent worktree (isolated branch) and tmux session—completely isolated.

Step 3: Monitoring Loop

Checks every 10 minutes via cron; auto-restarts on failure (max 3 retries). Instead of polling the Agent directly (too expensive), it reads a JSON registry to determine status:

Is tmux session alive?
Are there pending PRs?
CI status
Auto-restart on failure

Step 4: Create PR

Agent commits, pushes, and creates a PR via gh pr create.

Definition of Done (Critical):

PR created
Branch synced to main (no merge conflicts)
CI passed (lint, types, unit tests, E2E)
Codex review passed
Claude Code review passed
Screenshots included (if UI changes)

Step 5: Automated Code Review

Three AI models review, each with its own strength:

Review Model	Characteristics
Codex Reviewer	Edge case expert, most detailed, logic errors, race conditions
Gemini Code Assist	Free, excels at security and scalability issues
Claude Code	More conservative, suggestions often over-engineered

Step 6: Automated Testing

The CI pipeline runs extensive automated tests:

Lint and TypeScript checks
Unit tests
E2E tests
Playwright tests (against preview environment)

New Rule: Any UI change must include a screenshot, or CI fails. This significantly reduces review time.

Step 7: Human Review

Receive Telegram notification: “PR #341 ready for review”

At this point, the following is done:

CI passed
Three AI reviews passed
Screenshots show UI changes

Review takes only 5-10 minutes; for many PRs, reading code isn’t even necessary—just looking at the screenshots is enough.

Step 8: Merge

PR is merged. A daily cron job cleans up orphaned worktrees and the task registry.

Ralph Loop V2

This is an upgrade of the Ralph Loop. Unlike the traditional Ralph Loop which runs the same prompt every time, this system is different:

When an Agent fails, Zoe doesn’t retry with the same prompt. Instead, she adapts:

Agent context exhausted? → “Focus only on these three files”
Agent went in the wrong direction? → “Stop, the client wants X not Y; here is what they said in the meeting”
Agent needs clarification? → “Here is the client’s email and what their company does”

Zoe also proactively finds work to do:

Morning: Scans Sentry → Finds 4 new errors → Spawns 4 Agents to investigate and fix
After Meetings: Scans meeting notes → Marks 3 feature requests mentioned by clients → Spawns 3 Codex Agents
Evening: Scans git log → Spawns Claude Code to update changelog and client documentation

Agent Selection Guide

Agent	Best For
Codex	Backend logic, complex bugs, multi-file refactoring, cross-repo reasoning (90% of tasks)
Claude Code	Faster on frontend work, fewer git operation permission issues
Gemini	Strong design sense, generates HTML/CSS specs for Claude Code to implement

Zoe’s Selection Logic: Billing system bug → Codex; Button style fix → Claude Code; New dashboard design → Gemini designs first, Claude Code implements.

Current Bottleneck

RAM — Each Agent requires an independent worktree and node_modules. A Mac Mini with 16GB can only run 4-5 Agents.

The author purchased a Mac Studio M4 Max with 128GB RAM ($3,500) to scale.

Future Outlook

By 2026, there will be an explosion of one-person million-dollar companies. The next generation of founders won’t hire a team of 10, but will handle it solo using the right system:

An AI orchestrator acts as your extension (like Zoe), delegating to Agents specialized in different business functions: engineering, customer support, operations, and marketing.

Summary

The core insights of this system:

Two-Layer Architecture: Orchestration layer (OpenClaw) handles business context, Agent layer handles code execution
Specialized Context: Specialization is achieved not through different models, but through different contexts
Automation Level: From code generation to review to testing, almost fully automated
Human-AI Collaboration: Humans only make high-value decisions (merging PRs), everything else is automated

Cost: Claude ~ $100/mo, Codex ~$ 90/mo, can get started for $20.

Want to set this up? Copy the original text into OpenClaw and tell it: “Implement this agent swarm setup for my codebase.” Done in 10 minutes.

Further Reading: