💃 The Next Wave of Technical Debt Talks Back 🕺

Good morning, Salesforce Nerds!

Every agent makes sense on its own. But, that’s the trap. 🪤

Service built a case triage agent. Sales built a prospect researcher. Operations built one that updates records while everyone sleeps. 🏆

Now ask three questions. How many agents run in your org, who owns each one, and what can they touch?

If you hesitated, you’re not alone.

A 2026 OutSystems survey of nearly 1,900 IT leaders found that 94% worry AI sprawl is raising complexity, technical debt, and security risk. And only 12% run a centralized platform to manage it. 😬

We’ve seen this movie.

Applications became shadow IT. Integrations became spaghetti. 🍝 Automation became seven overlapping flows on the Opportunity object.

Nobody dared touch any of them.

Agents are next, and they talk back. Let’s fix that in your org. 👇️

TABLE OF CONTENTS

💃 The Next Wave of Technical Debt Talks Back 🕺

• No Owner, Many Clones

• Agents Hoard Access

• Dependencies Nobody Drew

• Govern Before You Scale

• Guardrails Enable Speed

EVERYONE BUILT, NOBODY OWNS

NO OWNER, MANY CLONES

Sprawl starts with a reasonable sentence: “We’ll just build our own.”

Marketing builds an agent. Support builds another, and neither bothered to ask IT. 🎭

Fragmented ownership is the root cause. Departments pick different tools, data sources, and standards.

Nobody holds the pen on the whole picture.

Duplication is the symptom. Two agents answer the refund question from two separate instruction sets.

One says 30 days. One says 45. Both sound confident. 🤹

Every agent is an app with a personality and a permission set. Each carries its own instructions, subagents, and business rules, which makes it a small, unreviewed policy document with a voice. 📜

The cure is reuse. One governed “check entitlement” action, called by five agents, beats five homemade versions. 🧩

PERMISSIONS NEVER RETIRE

AGENTS HOARD ACCESS

Every agent runs as a user.

In Agentforce, that is a dedicated agent user, and its permission sets define what the agent can see and do. The platform supports least privilege, but nothing forces you to practice it. 🔑

Watch the creep. A builder grants one extra object so the demo works. Another adds edit access to fix a test failure, and nobody removes either.

Six months later, a “read-only research agent” is updating opportunities. Nobody remembers granting that access, and nobody can say why. 🕵️

The risk scales with autonomy. A person with too much access needs a bad day to cause damage. An agent needs only an ambiguous instruction.

Visibility lags behind. EY’s September 2026 survey found that 26% of organizations using agentic AI cannot detect unauthorized agents operating inside their walls. 🚨

Permission sets show what an agent can do. They do not record why it can, who approved it, or when to revisit. 🧾

GOOD LUCK DEBUGGING THAT

DEPENDENCIES NOBODY DREW

Agents call actions. Actions call Flows. Flows call Apex. Apex calls an API that another team owns.

Draw that on a whiteboard and watch the marker run dry. 🕸️

Then add handoffs. One agent escalates to another agent, a human, or a queue that moved last sprint.

When a customer gets a wrong answer, which link failed? 🔍

Change makes it worse. Someone edits a shared Flow for a good reason, three agents change behavior on Monday, and none of their owners get the memo. 🎲

Native tooling helps, within limits.

Agentforce Observability, formerly Command Center, rolls up activity, error rates, escalation frequency, and latency across production agents.

Testing Center runs synthetic interactions in a sandbox to check topic and action selection before you ship. 🧯

Both tools tell you how an agent behaves. Neither tells you whether the agent should exist, or what breaks when you change its neighbor.

That part is architecture. 🏗️

BUREAUCRACY, BUT MAKE IT USEFUL

GOVERN BEFORE YOU SCALE

Governance sounds like a meeting. But, done right, it works like a registry. 📋

Just four (4) artifacts do most of the work:

• Inventory: every agent gets a record, and an agent missing from the registry does not ship.

• Ownership: one named human per agent, plus a backup.

• Lifecycle: defined stages from pilot to retirement, with review dates someone reads.

• Reusable capabilities: a shared catalog of vetted actions, so teams compose instead of clone.

Custom Metadata Types make a cheap registry. It deploys through your normal pipeline and gets reviewed in a pull request. 🛠️

Agent_Registry__mdt agent = Agent_Registry__mdt.getInstance('Case_Triage_Agent');

String owner = agent.Owner__c; // one accountable human
String scope = agent.Data_Scope__c; // objects and fields in play
String permSet = agent.Permission_Set__c; // the agent user's access
Date nextReview = agent.Review_Date__c; // an access review, not a suggestion

A registry is a boring document that prevents exciting incidents. 🥱

Governance also pays back in speed. Teams that reuse vetted actions skip the rework, and reviewers stop relitigating the same permission request. ✅

THE QUESTION THAT MATTERS

GUARDRAILS ENABLE SPEED

Agentforce adoption is not a contest to build the most agents. Anyone can build an agent before lunch. 🚀

The real question is harder. Can your organization govern, maintain, secure, and evolve the agents it already runs?

Ask before the next sprint, not after the next incident. 🧭

Architecture governance does not block AI adoption. It keeps adoption alive after launch day. 🛣️

Guardrails do not slow the car. They let you take the curve at speed. 🏎️

Technical debt charges interest. Agent debt charges it in trust.

SOUL FOOD

Today’s Principle

and now....Salesforce Memes