Practical patterns, starters & CLI tools for loop engineering with AI coding agents. Design systems that prompt and orchestrate agents (inspired by Addy Osmani and Boris Cherny). Includes loop-audit, loop-init, loop-cost.
TypeScript agentic-aiai-agentsai-codinganthropicautomationclaudeclaude-codecodexcoding-agentsdevops-automationdevtoolsgithub-actionsgrokllmloop-engineeringmcpprompt-engineering
import { readFile, copyFile } from 'node:fs/promises';
import { createHash } from 'node:crypto';
import path from 'node:path';
import { runInSandbox } from '@cobusgreyling/loop-sandbox';
export async function runSwarm(root, command, args, options) {
const { count, ...sandboxOptions } = options;
console.log(`\n🐝 Launching loop-swarm with ${count} sequential agents (serialized to prevent worktree races)...`);
const results = [];
for (let i = 0; i < count; i++) {
console.log(`\n▶️ Agent ${i + 1}/${count}`);
results.push(await runInSandbox(root, command, args, sandboxOptions));
}
console.log(`\n🧠 Analyzing swarm results...`);
// Filter out runs that failed (non-zero exitCode)
const successfulRuns = results.filter(r => r.exitCode === 0);
const failedRuns = count - successfulRuns.length;
if (failedRuns > 0) {
console.log(`⚠️ ${failedRuns} agent(s) failed with non-zero exit codes. They are excluded from consensus voting.`);
}
// Count identical patches among successful runs
const hashCounts = {};
let noChangeCount = 0;
for (const r of successfulRuns) {
if (!r.hasChanges || !r.patchFile) {
noChangeCount++;
continue;
}
const buffer = await readFile(r.patchFile);
const hash = createHash('sha256').update(buffer).digest('hex');
if (!hashCounts[hash]) {
hashCounts[hash] = { count: 0, files: [] };
}
hashCounts[hash].count++;
hashCounts[hash].files.push(r.patchFile);
}
const threshold = Math.floor(count / 2) + 1; // Strict majority of total launched agents
if (noChangeCount >= threshold) {
console.log(`✅ Consensus reached! ${noChangeCount}/${count} agents produced NO changes (success).`);
return {
reached: true,
majorityHash: null,
majorityCount: noChangeCount,
totalPatches: successfulRuns.length - noChangeCount,
consensusPatchFile: null,
divergentPatches: []
};
}
// Find majority among patches
let maxCount = 0;
let majorityHash = null;
for (const [hash, data] of Object.entries(hashCounts)) {
if (data.count > maxCount) {
maxCount = data.count;
majorityHash = hash;
}
}
const divergentPatches = Object.values(hashCounts).flatMap(d => d.files);
if (maxCount >= threshold && majorityHash) {
console.log(`✅ Consensus reached! ${maxCount}/${count} agents produced the exact same patch.`);
const representativePatch = hashCounts[majorityHash].files[0];
const consensusPatchFile = path.join(path.dirname(representativePatch), 'consensus.patch');
await copyFile(representativePatch, consensusPatchFile);
console.log(`🎉 Consensus patch saved to: ${consensusPatchFile}`);
return {
reached: true,
majorityHash,
majorityCount: maxCount,
totalPatches: successfulRuns.length - noChangeCount,
consensusPatchFile,
divergentPatches: divergentPatches.filter(f => !hashCounts[majorityHash].files.includes(f))
};
}
const maxAgreement = Math.max(maxCount, noChangeCount);
console.log(`❌ Swarm failed to reach consensus. (Highest agreement: ${maxAgreement}/${count})`);
return {
reached: false,
majorityHash: null,
majorityCount: maxAgreement,
totalPatches: successfulRuns.length - noChangeCount,
consensusPatchFile: null,
divergentPatches
};
}