From 48-Hour Bottlenecks to 20-Minute Fixes: How Michelle Mayes’ Agentic Army Automates Campaign & Engineering Audits

For most people, “AI adoption” translates to generating basic templates, tools, or virtual assistants.

But for Michelle Mayes, AI is a full-blown infrastructure rather than a mere assistant.

With a background in training custom machine learning models dating back a decade, Michelle has transitioned to the era of LLMs without missing a beat. She currently acts as a builder and coach in her role as SVP of Data, AI & Product at Pushnami, designing systems to transform how the company writes code, analyzes data, and solves its engineering problems.

This Masterclass explores Michelle’s AI strategy for bypassing the IT queue and shrinking two-day tracking issues into a 20-minute task. Her story highlights what’s possible when you move past basic prompts and create an infrastructure that solves your team’s real-life problems.

The Death of the Support Ticket: Managing 25+ Coding Agents

In a traditional tech stack, a bug report or feature request follows a rigid path: log a ticket in Jira, wait for sprint planning, and hope it gets picked up next week. 

Michelle has entirely upended this workflow. 

She has roughly 25 autonomous coding agents running locally and working on engineering problems 24/7. Using advanced AI coding harnesses like Claude Code and Codex, she spins up automated sessions directly from her phone.

Instead of relying on basic, single-prompt chatbots, her setup relies on a master agent capable of spawning anywhere from one to eight sub-agents. These sub-agents work in tandem, delegating tasks and executing distinct software problems simultaneously.

To see how these sub-agents systematically eliminate the traditional support queue, look at how the daily engineering workflow shifts in practical terms:

Still, deploying an automated workforce of this scale goes beyond generating raw code. You also need a flawless, real-time context layer to keep the agents from making critical errors.

The Secret to High-Accuracy AI? Context Management

A common complaint among tech leaders is that AI creates "slop" or hallucinates incorrect answers. According to Michelle, the root cause is poor context management, not AI itself.

To build an army of accurate agents, Michelle maintains a massive Retrieval-Augmented Generation (RAG) database for Pushnami. 

This database serves as a centralized, living brain for her agents, making sure they pull from verified facts instead of guessing. Keeping this repository constantly updated is crucial to ensuring the agents achieve near-instantaneous accuracy rather than churning for 25 minutes trying to find a single correct data source. 

Here’s a quick snapshot of how Pushnami’s RAG architecture works:

• Automated Data Collation: The system automatically ingests and updates information multiple times a day from sources including: Slack, Confluence, internal emails, customer help centers, and GitHub repositories.
  ‍
• Hybrid Search & Relevancy Scores: When an agent queries the database, a hybrid search algorithm evaluates the metadata, identifies themes, and attaches a precise relevancy score so the agent pulls the exact data it needs.
  ‍
• Compounding Knowledge Loops: At the end of every active session, the agent updates the master documentation based on whether its final answer was correct. This means the agents steadily grow smarter over time.

Michelle's engineering catchphrase has become "slap an MCP on it." She relies heavily on Model Context Protocol (MCP) servers to bridge gaps between fragmented developer tools and LLMs. Her agents always have a unified view of the company's code and workflows as a result.

From 2 Days to 20 Minutes: Optimizing Everflow Data Streams

Automating software development is one thing, but forcing AI to audit complex marketing data is an entirely new challenge. 

Pushnami processes massive traffic volumes, routing every click and conversion stream through an Amazon SQS queue directly into Looker. To bypass the data bottleneck, Michelle designed a workflow that turns raw data streams into instant answers:

• The 50x50 Dimension Bottleneck: Troubleshooting redirect flows or sudden campaign drops used to be a slow, manual strain. Because a single user journey can pivot across a 50-by-50 grid of overlapping dimensions, a standard performance investigation historically consumed one to two full days of an analyst's time.
  
  
• The Everflow API Foundation: AI agents cannot solve high-dimensional data problems without structured guardrails. Pushnami used Everflow APIs to extract and map out data relationships, enriching their Looker models to create a clear structural blueprint of partners and postbacks.
  
  
• The 20-Minute Resolution: With the data layer pre-modeled, the operational bottleneck vanished. Investigations that previously drained 48 hours of manual labor and oversight are now systematically wrapped up in 20 minutes.

The speed of this setup played out live during this Masterclass interview. While our call was running, one of Michelle’s background agents quietly finalized an open campaign investigation. The agent instantly parsed the metadata, connected the tracking dimensions, and flagged the exact root cause.

And so instead of a data team losing a weekend to manual database queries, the engineer's only remaining task was a quick, definitive check to confirm the agent's findings.

Where the Agentic Army Meets Its Match

Even a highly sophisticated automated army has its operational limits. 

For Michelle, the toughest battle on the digital frontline isn't writing code. Instead, it’s conquering end-to-end testing and quality control (QA).

While AI agents excel at deploying code patches and isolating data anomalies, verifying that those changes function flawlessly across legacy frameworks remains a major technical hurdle. Michelle says you can get about “95% there” but there are still hallucinations to deal with.

To win this battle and eliminate manual QA bottlenecks, Michelle is actively building a self-healing, self-upgrading software pipeline. Under this framework, an agent will write a code patch and autonomously run an integration testing suite to check its own work. 

Rather than relying on easily hallucinated text logs, the agent will record a live video of itself navigating the user interface to prove the changes actually worked. Once that visual proof is compiled, the system will autonomously merge the change straight into production

Reaching this level of autonomy will require a few more model upgrades from major AI labs, but finding immediate value with automated workflows starts with a much simpler experiment. 

Overcoming the Noise: Michelle's Advice for AI Skeptics

You don’t need a massive corporate mandate or a dedicated data science department to start proving the business value of automation.

For marketing and tech leaders paralyzed by the AI hype cycle, moving from theory to practice is overwhelming. When your data team is still losing days to manual database queries and your engineering queue remains stalled by support tickets, abstract vision statements don't help. 

To move past basic prompts and understand what a digital workforce can actually handle, she recommends a simple, tactical framework: The 2-Hour Full-Task Experiment.

By tackling a highly complex problem right out of the gate rather than a trivial task, leaders can see past the surface-level noise and realize what AI architectures are truly capable of. Michelle has demonstrated firsthand that the fastest way to eliminate operational friction is to give your agents the trusted context they need to solve it.