Skip to Content
0%

Forming, Storming, Norming, Performing: Four Stages Every AI Agent Goes Through

AI agents go through the same four stages as new teams. Agent Optimizer helps you see the storm, find what's causing it, and prove the fix worked.

If you’ve ever joined a newly formed team, you know how it goes.

The first few weeks are polite. Everyone’s figuring out their role, the plan looks great on paper, and nobody has disagreed about anything yet. Then the first real project hits. Assumptions collide, things fall through the cracks, and you find out what the plan didn’t account for. If the team works through it, something changes. People learn how to work together, the friction fades, and eventually the team is doing its best work without anyone forcing it.

In 1965, psychologist Bruce Tuckman gave those stages names that stuck: forming, storming, norming, and performing. More than 60 years later, it’s still one of the best-known models for how teams grow. The model also fits something Tuckman never imagined: AI agents.

Your agents are part of your team now. They answer customers, resolve cases, qualify leads, and support employees, working next to the people they help. And like any new team member, an agent tends to go through four stages, as illustrated here.

Forming: The agent is built and tested. 
Storming: It meets real customers, and reality pushes back. 
Norming: You find what’s wrong and fix it. 
Performing: Improvements are proven, and they compound.
And then it starts again.

Most teams plan carefully for forming and hope they’ll skip storming. They rarely do. Every agent storms. When an agent struggles, the signs are often small. It might escalate a few more cases each week or give answers that are almost right. Often, nobody notices until the numbers slip.

Agent Optimizer is built to help you through the storms. It’s a conversation-first agent that helps you build and improve AI agents. It works with you through every stage, from the first line of instructions to the hundredth proven improvement. You talk to it the way you’d talk to a colleague.

Forming: Build it, then prove it’s ready

Forming is where your agent takes shape. Describe what you need: “An agent that helps travelers rebook a missed connection, checks fare rules, and hands off to a human for anything over $500.” Agent Optimizer drafts the agent in Agent Script, combining deterministic business rules with Atlas Reasoning Engine planning, and previews the result.

Because it builds on the Salesforce Platform, your agent doesn’t start from scratch. It works with the customer data, Flows, Apex, and permissions your company already runs on.

Testing is coming soon to Agent Optimizer. Before your agent meets a single customer, it will generate test conversations, and you decide what passing means. Tests are graded by your own custom scorers, so “good” reflects your policies, your tone, and your definition of a resolved issue. When the agent passes, it ships through the sandboxes, permissions, and release process your team already trusts.

Storming: See what’s happening in production

No test suite can predict everything real customers will do. Within days of launch, your agent is likely to see questions nobody planned for, a backend system that returns something unexpected, and an instruction that behaved in testing but not at scale.

Storming is normal. The danger is storming you can’t see.

Ask Agent Optimizer where your agent is struggling. It reads production sessions one by one, not just aggregate metrics, and ranks the recurring problems behind escalations, abandonment, and low satisfaction by impact. With session tracing turned on, it reads the full reasoning trace behind each response, along with what your agent said: which topic the agent chose, which actions it called, which knowledge it retrieved, and what it cited. A transcript tells you what happened. The trace tells you why.

For standard use cases, Agent Optimizer is included for every Agentforce customer at no extra cost, so you can watch your agents closely and still keep costs under control. 

Norming: Find the cause, then fix it 

Teams get through storming by working through the friction. The same goes for agents.

For each issue, Agent Optimizer finds the root cause, down to the exact step where things went wrong, and sorts it by type, such as a broken action, a knowledge gap, or an off-topic request, so it reaches the right owner.

Then it recommends a fix, such as a clearer instruction, a new topic, or a missing knowledge article. Agent Optimizer can build the fix (and test it, once testing arrives), and you choose how much it does on its own and where it stops for your team’s review.

Performing: Prove it works  

Every team has made a change that looked great on paper and fell flat in practice. High-performing teams don’t guess. They test.

A/B experimentation, coming soon to Agent Optimizer, lets you run a change against your current agent on real traffic before you roll it out to everyone. You judge the results with the same custom scorers you used in testing.

Proven improvements compound. Each fix stays in your agent’s configuration, so every new round of testing starts from the improved agent and builds on the last one.

And then it starts again                     

Teams don’t stay in performing forever. When something changes, like a new member, a new goal, or a new market, they often go back through the stages.

Agents do too. A new product launches, a policy changes, you add a topic or switch models, and your agent goes back to forming. The difference is what happens next. With each cycle through Agent Optimizer, the storm can get shorter, the cause clearer, and the fix faster to prove. Each lap takes your agent from novice toward expert.

Agent Optimizer works across Agentforce agents, including service, sales, employee-facing, and multi-agent teams.

Every agent storms. Agent Optimizer helps yours come out stronger after each one.

Agent Optimizer is available in beta, with testing and A/B experimentation on the way. To get started, contact your account executive and read the help documentation.

Any unreleased services or features referenced in this blog post are not currently available and may not be delivered on time or at all. Customers should make their purchase decisions based upon features that are currently available.

Building is the easy part

Every agent in production needs an operating model behind it. The Grown-Up’s Guide to Operating an Agentic Enterprise lays out the roles you need and the metrics to watch as your agents grow.

Get the latest articles in your inbox.

Get the latest articles in your inbox.