🚜 Deploying Heavy Machinery: Multi-Agent Swarms, Automated Triggers & Background Cron Workers

"Look at Slide 2 on your screen. How do we break a fragile 'Mega-Prompt' into a rock-solid Supervisor-Worker Crew?

The golden rule of multi-agent engineering is: One Agent = One Job = One Strict JSON Hand-off.

Look at the three subcontractors on Slide 2:
• Subcontractor #1 is the Extractor Agent (The Excavator): Its system prompt is ruthless. It has zero opinions. It takes a messy PDF, site photo, or CV and outputs pure structured JSON: {docId, vendor, lineItems[], confidenceScore}. It is forbidden from approving or summarizing.
• Subcontractor #2 is the Auditor Agent (The QS / QA Inspector): It never even sees the messy raw text! It only receives the clean JSON from Agent #1 and compares it against your Golden Reference Table — your FIDIC clauses, DOSH safety rules, or HR rubric — and outputs a deterministic verdict: {status: 'PASS' | 'FLAGGED', violations[], severity}.
• Subcontractor #3 is the Dispatcher Agent (The Site Clerk): It takes the verified verdict from Agent #2, writes the database record, and drafts the notification card for the human engineer."

"On a deep excavation site, nobody stands by the sump pit at 3:00 AM flipping the dewatering pump switch by hand. You have a float switch (an event trigger) and a scheduled pump cycle (a timer trigger).

Your AI fleet needs the exact same two switches so it works while you sleep:

⏰ Pattern A: Time-Driven Cron (The Morning Shift Briefing):
Using Cloud Scheduler, GitHub Actions Cron, or Node-Cron (`0 6 * * 1-5` — 6:00 AM Monday to Friday), your background worker wakes up before site engineers even finish their teh tarik. It queries all unreviewed logs from the last 24 hours, runs the 3-agent pipeline in batch, and pins the Morning Executive Digest at the top of your dashboard by 06:15 AM.

⚡ Pattern B: Event-Driven Webhook (The Float Switch):
The moment a subcontractor uploads a file to Cloud Storage, submits a form, or a Firestore/SQL row is inserted, an HTTP `POST /api/webhook/new-submission` fires immediately. Your server returns `202 Accepted` in 50 milliseconds so the user's screen never freezes, while the 3-agent crew processes the document asynchronously in the background."

"Now pause and think like a Chief Engineer. The moment you connect an autonomous agent crew to a cron timer or a webhook, what is the biggest danger? Runaway cranes!

If a user double-clicks 'Submit', or a webhook retries 5 times, or two agents argue in a `while(true)` loop, you can burn through your Gemini API quota in 90 seconds or insert 10 duplicate rows into your database.

Every production multi-agent app at Gamuda MUST install these three Tower Crane Anti-Collision Lockouts:

🔒 Lockout #1: Idempotency Keys (No Double-Piling):
Before your Supervisor runs, hash the input (`SHA-256(doc_id + file_version)`). Check your database first: if that exact hash already has a `COMPLETED` or `IN_PROGRESS` status, return the cached result immediately. Never process the same PDF twice!

🚦 Lockout #2: Concurrency Semaphore (`MAX_CONCURRENT = 3`):
If 50 site reports arrive at 5:00 PM, do not fire `Promise.all()` on 50 parallel LLM calls — you will immediately hit `429 Too Many Requests`. Use a queue with `MAX_CONCURRENT = 3` and exponential backoff.

🛑 Lockout #3: Hard Circuit Breaker (`MAX_STEPS = 5`):
Never write an unbounded retry loop. Cap every agent run at `MAX_STEPS = 5` and a hard `25-second timeout`. If the agent still can't validate the schema on attempt 3, abort cleanly and route to human review."

"Even the most advanced autonomous TBM stops immediately when ground-penetrating radar detects an unmapped high-voltage cable or gas main. In civil engineering, we call this a Mandatory Inspection Hold Point (Request for Inspection / RFI).

In enterprise AI, autonomy does not mean letting an agent approve a RM 250,000 Variation Order or reject a candidate without human oversight. Instead, we design a 3-Tier Traffic Light Hold Point:

• 🟢 Green Zone (Auto-Commit — 80% of routine volume): Confidence score ≥ 92%, zero safety or commercial discrepancies, value under RM 5,000. The Dispatcher Agent logs it automatically and marks it `AUTO_VERIFIED`.
• 🟡 Amber Zone (Hold for Review — `PENDING_ENGINEER_APPROVAL`): Confidence between 70%–91%, or minor commercial variance. The agent prepares the entire evidence package — highlighting the exact clause and extracted numbers side-by-side — and pauses in an Approval Queue with 1-click [✅ Approve] / [❌ Reject] buttons for the human engineer.
• 🔴 Red Zone (Immediate Halt): Safety critical violation, missing mandatory signature, or prompt-injection attempt. The pipeline halts and escalates immediately."

"Alright team, let's lock in our mission for Sprint 4! Notice on the calendar that because of the Deepavali festive week buffer, our next live session — Sync #5 (Production Hardening, Security & Load Testing) — takes place on Tuesday, 17 November 2026 at 09:00 AM MYT.

Before we meet for Sync #5, every pod needs to ship three upgrades to your live deployed application:

1️⃣ Checklist #1 — Specialist Split: Refactor at least one core workflow from a single mega-prompt into a 2-to-3 Agent Supervisor Pipeline (e.g., Extractor → Auditor → Dispatcher) passing strict JSON between steps, with visible step-by-step status badges in your UI so users can see each specialist agent working.
2️⃣ Checklist #2 — Automated Trigger + Idempotency: Add a scheduled cron route or event webhook (plus a 'Simulate Night-Shift Run' button on your dashboard) protected by an idempotency check and `MAX_STEPS` guardrail.
3️⃣ Checklist #3 — Human-in-the-Loop (HITL) Approval Queue: Add an Amber-Zone status (`PENDING_REVIEW`) where flagged items wait for a human clicks Approve or Reject."