Top 6 Agent Principles | Brothers Automate
Talk to the Brothers Book Call
BROTHERS_AUTOMATE / RESOURCES · AGENT PRINCIPLES ● BUILD GUIDE · 06 PRINCIPLES
HOW WE BUILD · OPERATOR PLAYBOOK

Top 6 Agent Principles.

The rules we follow every time we build an AI agent or workflow for a service business. No exceptions.

PRINCIPLE_01 MOST CRITICAL

Graceful error handling

WHAT WE KNOW

We build for failure, not just the happy path. APIs fail, return empty, or time out. Without handling this, agents loop infinitely or silently die mid-workflow.

EXAMPLE SCENARIO

Agent calls a CRM, gets null back, tries to access contact.email, and crashes the entire run.

HOW TO IMPLEMENT 
if (!contact || !contact.email) {
  // log · notify · fallback
  // NEVER continue blindly
  return fallback_action()
}
PRINCIPLE_02 SAFETY

Human-in-the-loop checkpoints

WHAT WE KNOW

We know which actions to automate and which to gate. Some are irreversible, sending emails, deleting records, charging payments. The wrong automation at the wrong moment is worse than none.

EXAMPLE SCENARIO

Agent drafts and sends a refund email to 200 customers before anyone reviews. One line was wrong. You can't unsend it.

HOW TO IMPLEMENT 
⚠️ NEVER execute without approval:
- send_email
- delete_record
- submit_payment

Draft → show human → confirm → act.
PRINCIPLE_03 ARCHITECTURE

Lean prompt + on-demand skills

WHAT WE KNOW

We keep system prompts lean, identity, tone, guardrails only. Process logic lives in separate skills loaded on demand. Leaner prompts mean more predictable, more maintainable agents.

EXAMPLE SCENARIO

A 3,000-word system prompt covering 12 workflows. The agent starts mixing up instructions between processes.

HOW TO IMPLEMENT 
System prompt: identity + rules only
  → under 300 words

Skills (loaded on demand):
  handle_refund.md
  qualify_lead.md
  draft_proposal.md
PRINCIPLE_04 MAINTAINABILITY

Keep data external and editable

WHAT WE KNOW

We separate instructions from data. Pricing, FAQs, and templates live in external sources the agent reads live, so anyone on the team can update content without touching the agent.

EXAMPLE SCENARIO

Pricing changes. Agent still quotes last month's rates because they're hardcoded in a prompt only a developer can edit.

HOW TO IMPLEMENT 
"Before quoting pricing, fetch the
current rate card from Notion page
{pricing_page_id}."

// Team updates Notion.
// Agent always reads fresh data.
PRINCIPLE_05 RELIABILITY

Idempotency, design for retries

WHAT WE KNOW

We design every workflow to be safe to re-run. Triggers fire twice. Retries happen. An idempotent workflow produces the same result regardless of how many times it executes.

EXAMPLE SCENARIO

A Zap fires twice after a timeout. Agent creates two CRM contacts, sends two welcome emails, charges the card twice.

HOW TO IMPLEMENT 
existing = search_crm(email)

if existing:
  update_record(existing.id)
else:
  create_record(new_data)

# Never blindly CREATE, check first.
PRINCIPLE_06 OBSERVABILITY

Log what the agent did and why

WHAT WE KNOW

We build observability in from the start. Every meaningful action, what the agent read, decided, and executed, is logged. That's how you debug, audit, and improve over time.

EXAMPLE SCENARIO

Agent misroutes 40 support tickets to the wrong team. No log exists. You can't diagnose why or prove it happened.

HOW TO IMPLEMENT 
After each action, log:
  - input received
  - decision made
  - action taken
  - timestamp + run_id

Store in: {audit_log_table}
BROTHERS_AUTOMATE / SECTION 06 · TALK TO THE BROTHERS 30-MIN DISCOVERY CALL · NO PITCH
SCHEDULE_INTAKE · 30_MIN ● ACCEPTING NEW BUILDS
GET ON OUR CALENDAR

Tell us what's eating your time.
We'll show you what we'd build to fix it.

30 minutes. We listen, we map your bottleneck, we tell you straight whether we can help and what it'd cost. No deck. No pressure. Either you walk away with a plan or you walk away with clarity.

Talk to the Brothers OR SEE THE BUILDS → OR SEND A NOTE →
FREE CONSULTATION · NO OBLIGATION · CUSTOM-BUILT FOR YOUR BUSINESS