Home

Blog

How Persefoni Built a Council of 12 AI Agents to Build on Workato

AI Agents & Orchestration

Customer Stories

Persefoni’s Matt Palmer showed how a council of specialized AI agents builds in Workato, with a human still deciding what ships.

Hannah Peacock

Director, Solutions Marketing

Oct 14, 2026

Matt Palmer’s team of specialized agents takes a Jira ticket from backlog to built, tested, and reviewed in minutes. A human still decides what ships.

At World of Workato (WOW) in Las Vegas, Matt Palmer, Automation and Integration Lead at Persefoni, a carbon management and accounting platform, walked through the team of AI agents he used to build in Workato. He calls it the “Jedi Council.” Viet Huynh from Workato’s AI business solutions team introduced him.

Matt’s message was practical: split the work into narrow roles, give each agent only the tools it needs, and keep a human at the moves that matter.

One agent with all the knowledge kept making mistakes

Matt started in December with a single Claude chat session and a developer MCP he built on Workato’s APIs. He drove the architecture, the recipe build, review, and testing himself. A single autonomous agent was better, and he gave it skills describing what a recipe, a function, or a Salesforce trigger looks like.

It still made errors it had the context to avoid. When he asked why, the agent agreed it should have called the right skill. His conclusion: “It was too much knowledge for any one agent.”

So he split the knowledge and responsibilities across agents, each with a narrow focus. That, he said, is where the real improvement began.

Organize agents the way you would organize a team

Matt’s council has four groups:

Orchestration

“Yoda,” the orchestrator, picks up work from Jira and assigns it. The architect writes and updates Jira tickets to shape how work gets done, but doesn’t assign work.

Builders

Specialists for agentic skills, the connector SDK, recipes, and API endpoints.

Reviewers and testers

“Obi-Wan” is a pair-programming partner, “Qui-Gon” reviews, and “Ahsoka” tests. “Palpatine”, the one Sith, acts as an adversarial agent that pushes the others to question their work.

Foundation

A knowledge agent compiles what the agents learn and keeps their memories clean. A workspace steward creates the data tables, properties, folders, projects, and MCPs in the Workato workspace.

These are “Hermes” agents, not Workato Genies. Each runs on its own virtual machine with its own Slack identity, and most of their coordination happens in Slack. They began improving noticeably, Matt said, after he added an open source memory project to give each agent more room to remember.

Use the CLI for code and AIRO for the rest

Agents use the Workato CLI to pull recipe code to a local workspace, where they edit it and push it back. Matt likes having the code local: agents can review the whole workspace at once and answer questions like “find all my recipes that have this action.” He recently added the AIRO MCP for what the CLI can’t do, such as adding records to data tables or setting project and environment properties.

A live demo caught a silent failure

In the demo, “Yoda” picked up a real ticket from Matt’s backlog without being prompted. A builder started on a skill that searches Salesforce quotes. “Obi-Wan” flagged that a shortcut could return nothing and asked about the fallback.

“Ahsoka” then paused for human sign-off before executing the code. Matt approved, and she ran a live test against real Salesforce data in his sandbox. She found that searching only by approved status reported success while missing the account or opportunity ID it needed. It looked like it worked, but it didn’t. That was the risk “Obi-Wan” had flagged. “Qui-Gon” then read the code line by line, including the SOQL queries.

The point, Matt said, is not replacing builders:

“This isn’t about replacing your builders, it’s about giving every one of them a tireless apprentice who takes the first pass, and so your people spend their hours on judgment, not on building the plumbing.” — Matt Palmer, Automation and Integration Lead, Persefoni

“This isn’t about replacing your builders, it’s about giving every one of them a tireless apprentice who takes the first pass, and so your people spend their hours on judgment, not on building the plumbing.” — Matt Palmer, Automation and Integration Lead, Persefoni

Humans stay at the gates, and permissions are split

Agents don’t push to production. Matt does. He splits permissions the way you would for human staff: builders can write code in Workato, while reviewers and testers can’t. Only the orchestrator and architect can create Jira tickets. If a system has only a production instance, he wouldn’t let agents push there at all.

He is candid about limits. Observability is mostly Slack plus Workato logs of MCP calls. The CLI is more of a black box, and he can’t see in Workato which agent made a given API call. When he first put all 12 agents in one channel without ground rules, a single greeting set off about 2,500 messages in a thread, and he used “Hermes” slash commands to stop them.

The output gap, and where to start

Matt worked 397 Jira tickets in 2025. He says he is now close to four times that with three months of the year left, and that his output jumped once the full council was running in April. His role has changed, too:

“[With these agents], I feel like a manager with employees producing the work.” — Matt Palmer, Automation and Integration Lead, Persefoni

“[With these agents], I feel like a manager with employees producing the work.” — Matt Palmer, Automation and Integration Lead, Persefoni

For anyone who wants to try it, he suggests starting where he did: get comfortable in Claude Code, add the AIRO MCP, and install the Workato CLI. As you learn the rhythm, you can run additional sessions while a builder works. His limit is how much context he can juggle across sessions.

He hopes better testing features will eventually let him step out of the loop. For now, he’d put new teammates through the foundations first.

The future of work is being planned at World of Workato. Save the date for WOW 2027.