Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Fintech Brex is betting that the future of enterprise AI lies not in better orchestration, but in less.
As generative AI agents move from co-pilot to autonomous system, Brex CTO James Reggio says traditional agent orchestration frameworks become a constraint rather than an enabler. Instead of relying on a central coordinator or rigid workflows, Brex has built what it calls an “Agent Mesh”: a tight network of role-specific agents that communicate in simple language and operate independently, but with full visibility.
“Our goal is to use AI to actually make Brex disappear,” Reggio told VentureBeat. “We are aiming for total automation.”
Brex has learned that to achieve its goals, agents must work in narrow and specific roles to be more modular, flexible and auditable.
Reggio said the architectural goal is to allow every manager in a company “to have a single point of contact within Brex that manages all of their responsibilities, whether it’s managing expenses, travel requests, or approving spending limit requests.”
The journey from Brex Assistant
The financial services industry has long embraced AI and machine learning to manage the massive amounts of data it processes. But when it comes to bringing AI models and agents, the the industry took a more cautious path at first. Now, more and more financial services companies, including Brex, have launched AI-based platforms And multiple agent workflows.
Brex’s first foray into generative AI was with its Brex Assistant, released in 2023, which helped customers automate certain financial and spending tasks. It provides suggestions for completing expenses, auto-populates information, and tracks expenses that violate policies.
Reggio acknowledges that Brex Assistant works, but it’s not enough. “I think to some extent it’s still a technology that we don’t fully understand the limitations of," he said. "There are a lot of models that need to exist around this that are sort of developed by the industry as the technology matures and more companies build with it."
Brex Assistant uses several models, including Claude d’Anthropice and custom Brex models, as well as the OpenAI API. The assistant automates certain tasks, but remains limited in its level of contact.
Reggio said Brex Assistant still plays an important role in the company’s autonomy journey, primarily because its Agent Mesh product is integrated into the app.
Agent Mesh to replace orchestration
The consensus in the industry is that multi-agent ecosystems, in which agents communicate to accomplish tasks, require an orchestration framework to guide them.
Reggio, on the other hand, has a different view. "The deterministic orchestration infrastructure…was a solution to the problems we saw two years ago, which was that agents, like models, hallucinate a lot,” Reggio said. “They’re not very good with multiple tools, so you have to give them those degrees of freedom, but in a more structured and rigid system. But as the models get better, I think it starts to put a damper on the expanding range of possibilities.
More traditional agent orchestration architectures focus on either a single agent that does everything, or, more commonly, a coordinator/orchestrator and tool agents that explicitly define workflows. Reggio said both frameworks are too rigid and address problems more commonly found in traditional software than in AI.
According to Reggio, the difference is structural:
-
Traditional orchestration: predefined workflows, central coordinator, deterministic paths
-
Agent Network: event agents, specialized in roles, message-based coordination
Agent Mesh relies on the creation of networks of many small agents, each specialized in a single task. Agents, once again using the hybrid mix of models as with the Brex Assistant, communicate with other agents “in plain English” via a shared message stream. A routing model quickly determines which tools to invoke, he said.
A single reimbursement request triggers multiple tasks: a compliance check to align with spending policies, budget validation, receipt matching, and then payment initiation. While an agent can certainly be coded to do all of this, this method is “fragile and error-prone” and responds anyway to new information shared via a message stream.
Reggio said the idea was to disambiguate all of these separate tasks and put them in the hands of smaller officers. He likened the architecture to a Wi-Fi mesh, in which no single node controls the system: reliability emerges from many small, overlapping contributors.
“We found the idea of embodying specific roles as agents on top of the best platform to handle specific responsibilities fit very well, much like how you might delegate accounts payable to one team rather than expense management to another team,” Reggio said.
Brex defines three main ideas in the Agent Mesh architecture:
-
Config, where agent, model, tools, and subscription definitions are live
-
MessageStream, a log of every message, tool call and state transition
-
Clock, which guarantees a deterministic order
Brex has also built assessments into the system, in which the LLM acts as a judge and an auditing officer reviews each officer’s decisions to ensure they meet accuracy and behavioral policies.
Success so far
Brex says it has seen substantial efficiencies among its customers in its AI ecosystem. Brex did not provide third-party benchmarks or customer-specific data to validate these gains.
But Reggio said enterprise customers using Brex Assistant and the company’s machine learning systems “are able to achieve 99% automation, especially for customers who have really moved toward AI.”
This is a marked improvement from the 60-70% of Brex customers who were able to automate their spending processes before the launch of Brex Assistant.
The company is still at the beginning of its journey toward self-sufficiency, Reggio said. But if the Agent Mesh approach works, the most successful result might be invisible: employees no longer think about expenses at all.
