Inside Brex’s Agent Mesh: A New Architecture for Autonomous Finance

Brex is advancing a contrarian view in the rapidly evolving world of enterprise AI: as agents grow more capable, the industry may need less orchestration, not more. The company is building what it calls an “Agent Mesh” for autonomous finance — a distributed network of narrow, role-specific agents that communicate in plain language, operate independently, and are fully auditable.

The aim, according to Brex CTO James Reggio, is not simply to augment finance teams, but to automate away much of their work entirely. “Our goal is to use AI to make Brex effectively disappear,” he said. “We’re aiming for total automation.”

From copilots to autonomy in enterprise finance

The financial services industry has used AI and machine learning for years to manage risk, fraud, and large-scale data processing. When generative AI arrived, however, the sector moved cautiously. Early pilots focused on constrained use cases and human-in-the-loop review as firms weighed regulatory, compliance, and reputational risks.

That stance is shifting. More institutions — from large banks to fintech platforms — have launched AI-powered platforms and agentic workflows to handle tasks such as document processing, customer support, and transaction monitoring. Brex is part of this wave, but its long-term ambition goes beyond productivity gains: it wants to make key finance operations close to fully autonomous for customers.

The company’s first major step into generative AI was Brex Assistant, launched in 2023. The assistant helps automate finance and expense processes by suggesting how to complete expense reports, auto-filling missing information, and following up on policy violations. It taps multiple models, including Anthropic’s Claude, in-house Brex models, and OpenAI’s API.

Brex Assistant works, Reggio said, but it exposed both the promise and the current limits of generative AI in production finance workflows. The assistant can automate a meaningful portion of tasks, yet it remains a tool users interact with explicitly rather than a system that quietly handles work end to end.

Reggio noted that generative AI itself is still maturing, and the industry is collectively discovering patterns for safe, scalable deployment. “There’s quite a large number of patterns that need to exist around it that are kind of being developed by the industry as the technology matures and as more companies build with it,” he said.

Brex’s autonomy roadmap now centers on reducing the need for users to think about the system at all. The Agent Mesh is designed to feed into products like Brex Assistant but shift more of the work from interactive copilots to autonomous, behind-the-scenes agents.

Why Brex is pushing back on traditional orchestration

As multi-agent AI systems spread, a rough industry consensus has formed: to make agents reliable, they need orchestration. In most architectures, that means either a single, powerful agent coordinating everything or a central orchestrator managing specialized tool agents through predefined workflows.

Reggio argues that this assumption is becoming outdated as models improve. Deterministic orchestration frameworks, he said, were originally designed to compensate for early agent weaknesses: frequent hallucinations, difficulty juggling multiple tools, and unreliable decision-making unless tightly constrained.

Those constraints resulted in architectures that look more like classic enterprise software than adaptive AI systems — with strict workflows, rigid state machines, and a central coordinator deciding what happens when. That approach, Reggio contends, solves yesterday’s problems at the cost of tomorrow’s flexibility.

He contrasts two structural models:

• Traditional orchestration: Predefined workflows, a central coordinator, and deterministic paths that agents follow.
• Agent Mesh: An event-driven system composed of role-specialized agents that coordinate via messages rather than a central workflow engine.

In the traditional model, the orchestrator is responsible for sequencing tools and agents in a prescribed order. This can be easier to reason about and test, but it makes adaptation costly: changing policies, adding data sources, or introducing new capabilities often requires reworking the orchestration logic itself.

Brex’s view is that, as models become less brittle, the orchestration layer should do less directing and more enabling. In other words, instead of meticulously choreographing agent behavior from above, the platform should allow many small agents to react to events and messages, each within a narrow, well-understood scope.

Agent Mesh: a distributed, role-based agent network

Brex’s Agent Mesh replaces tightly scripted workflows with a network of narrow agents, each assigned a specific role and responsibility. These agents use the same hybrid model strategy as Brex Assistant — relying on third-party and in-house models — but are organized to work like specialized team members rather than a single monolithic assistant.

The core idea is to decompose finance processes into discrete, auditable steps and assign each step to its own agent. Agents communicate “in plain English” over a shared message stream, rather than through rigid function calls embedded in an orchestration tree. A routing model determines which tools to invoke as messages arrive.

A simple reimbursement request illustrates the difference. In a traditional orchestration system, an agent or orchestrator might handle the entire flow: check the expense against policy, validate budgets, match receipts, and trigger payment. All logic is centralized, and new constraints or features must be wired into that one flow.

Brex instead separates these concerns:

• A compliance agent checks that the request aligns with expense policies.
• A budget agent validates that funds are available.
• A receipt-matching agent ensures documents line up with the claim.
• A payment agent initiates the transaction when prerequisites are met.

Each agent listens to the shared message stream, reacts when relevant events appear, and then emits new messages that other agents can act on. In this design, no single agent is responsible for the entire end-to-end workflow, and no central coordinator dictates the sequence.

Reggio likens this to a Wi-Fi mesh network: there is no single node in control, but reliability emerges from many overlapping contributors. In organizational terms, Brex conceptualizes each agent as analogous to a specialized finance team — accounts payable, expense management, approvals — each with a clearly defined mandate.

This modularity is not just about flexibility; it also supports the company’s goal of giving every enterprise manager a single, unified point of contact into Brex. Behind the scenes, many agents may engage, but from the user’s perspective, requests related to spend management, travel, or approvals are handled seamlessly by what appears to be one interface.

Inside the Agent Mesh architecture: Config, MessageStream, and Clock

To make a distributed agent network reliable enough for finance, Brex defines three core architectural constructs within Agent Mesh:

• Config: This is where the definitions of each agent live — including which model it uses, what tools it can access, and what subscriptions or events it listens to. Config effectively encodes the roles and capabilities of the mesh.
• MessageStream: A log of every message, tool call, and state transition that flows through the system. This forms the backbone of observability: it provides a complete record of what each agent saw, decided, and did.
• Clock: A mechanism to ensure deterministic ordering of events. Even in an event-driven, distributed system, finance workflows must be replayable and explainable; the Clock component ensures that the sequence of actions can be reconstructed consistently.

These elements together turn a loosely coupled agent network into something that can meet the operational and compliance expectations of enterprise finance. Rather than relying purely on external observability tools, Brex builds evaluation and auditing directly into the fabric of the system.

One notable design choice is the use of large language models as evaluators. Within Agent Mesh, an LLM acts as a judge over agent decisions, providing an assessment layer that can score or flag outcomes. On top of that, an audit agent reviews decisions to confirm they align with accuracy standards and behavioral policies.

This layered evaluation does not eliminate the possibility of error, but it provides a built-in feedback loop. Because all messages and tool invocations are logged in MessageStream and ordered via Clock, audits can trace exactly how a particular decision was made and which agents contributed to it.

The result is a system where autonomy and accountability are designed to coexist: agents are free to act within their roles, but their actions are continuously recorded and subject to automated review.

What Agent Mesh means for AI architects and fintech leaders

For enterprise AI architects and fintech product leaders, Brex’s approach highlights several architectural and organizational implications as generative AI moves from pilots to production.

1. Narrow agents over generalists. Brex’s experience suggests that agents constrained to specific, well-defined roles are easier to reason about, test, and audit. This runs counter to the temptation to build a single, powerful generalist agent per domain. In regulated spaces like finance, role specialization can translate directly into clearer controls and more transparent decision trails.

2. Event-driven coordination instead of rigid workflows. By leaning on event streams and message-based coordination, Agent Mesh prioritizes adaptability. New policies, tools, or data sources can potentially be integrated by introducing or modifying agents subscribed to certain events, rather than re-authoring entire orchestration graphs.

3. Built-in observability as a first-class concern. Components like MessageStream and Clock emphasize that observability and determinism are not add-ons in an agentic system; they are foundational. For leaders designing AI platforms that must withstand audits or regulatory scrutiny, a persistent, queryable event log can be as critical as the models themselves.

4. Human experience as the unifying layer. Despite its internal complexity, Brex frames the goal in simple terms for end users: every manager should feel like they have a single point of contact handling spend, travel, and approvals. Architecturally, this means abstracting away the mesh and presenting a coherent interface while still allowing internal specialization.

Brex is not claiming that orchestration is obsolete; rather, it is arguing that the industry’s default pattern — a central, deterministic orchestrator — may need to give way to more distributed, event-driven designs as models improve. For organizations now considering their second or third generation of AI architecture, Agent Mesh offers one blueprint for that transition.

Early results and the road to ‘invisible’ expenses

Brex reports that its AI ecosystem has already delivered meaningful efficiency gains for customers. According to Reggio, enterprise customers using Brex Assistant and the company’s machine learning systems “are able to achieve 99% automation, especially for customers that really leaned into AI.”

Before Brex Assistant, customers were typically automating about 60% to 70% of their expense workflows. The company did not provide third-party benchmarks or customer-specific data to independently validate the 99% figure, so these results should be viewed as internal metrics rather than industry-wide proof points.

Still, the improvement illustrates the trajectory Brex is targeting: from partial automation with manual checks to near-complete automation for customers willing to adopt AI-centric processes. Agent Mesh is intended to push that trajectory further by shifting more work out of interactive assistants and into autonomous agents that operate continuously in the background.

Reggio emphasizes that Brex is still in the early stages of this autonomy journey. The Agent Mesh architecture is designed to scale, but its full impact will only become clear as more customers rely on it for core finance operations and as the underlying models continue to advance.

If the approach succeeds, Brex expects the most tangible outcome to be something users barely notice: expense management that effectively disappears from employees’ day-to-day concerns. Requests, approvals, compliance checks, and reimbursements would still be happening — but orchestrated by a mesh of specialized agents rather than by human coordinators or a visible central AI assistant.

For AI and fintech leaders, the question is whether this style of distributed agent architecture can generalize beyond Brex’s domain. What is clear from Brex’s experience so far is that moving from copilots to autonomy requires rethinking not only models and tools, but also the orchestration patterns that tie them together.