According to VentureBeat, Brex CTO James Reggio argues that traditional AI agent orchestration is now a constraint, not an enabler. The fintech’s solution is an “Agent Mesh,” a network of narrow, role-specific agents that communicate in plain language without a central coordinator. This architecture aims for “total automation,” with the goal of making Brex itself effectively disappear for enterprise managers handling spend or travel. The company’s earlier product, Brex Assistant launched in 2023, helped automate tasks but was limited. Now, Brex says customers using its AI and machine learning systems can achieve 99% automation, a major jump from the 60-70% possible before Brex Assistant. The company is still early in this autonomy journey but believes its mesh approach is the future.
Orchestration is the bottleneck
Here’s the thing: the entire industry has been building towards more sophisticated orchestration for AI agents. You have a coordinator agent that breaks down a task, farms out steps to specialist agents, and stitches it all back together. It makes sense, right? It’s how you manage complexity and hallucinations. But Brex is basically saying that whole paradigm is already legacy tech. James Reggio’s argument is that as the underlying models get vastly better, that rigid, deterministic orchestration framework starts to hold you back. It solves yesterday’s problem—unreliable agents—but creates tomorrow’s problem: inflexibility. So they’re ripping out the central brain and betting on emergent intelligence from a swarm.
How the Agent Mesh works
Instead of one smart agent or a boss-agent with worker drones, the Agent Mesh is a peer network. Think of it like a team of human specialists who all speak the same language and have a shared inbox. A reimbursement request comes in as a “message” to the stream. One agent, specialized in compliance, picks it up, checks it against policies, and posts its result. A budget agent sees that, validates funds, and posts. A receipt-matching agent does its thing. There’s no pre-defined workflow saying “do A, then B, then C.” The work gets done because specialized agents are listening for events they care about and chiming in. Reggio likens it to a Wi-Fi mesh—no single point of failure, and reliability emerges from the crowd. They even use a hybrid model mix, including Anthropic’s Claude and OpenAI’s API, and Claude Code reportedly helped them write 80% of this new codebase. That’s a huge deal for development speed.
The invisible enterprise
The real endgame here is fascinating. Brex isn’t just trying to make a better dashboard or a smarter chatbot. The goal is to make the entire financial operations layer of a company *disappear*. Reggio talks about a manager having a “single point of contact” that handles *everything*—but that point of contact isn’t a person, or even a single AI. It’s the entire mesh, working so seamlessly you stop thinking about the process entirely. Expenses get submitted, approved, reimbursed, and logged without you lifting a finger. The 99% automation claim is staggering if true. But look, they haven’t provided third-party benchmarks, so a healthy dose of skepticism is warranted. Still, moving from 70% to 99% automation isn’t just an incremental gain. It’s a phase change. It means you stop “doing expenses” and start just… spending money, with the system handling the rest invisibly.
A bet on simplicity
This feels like a classic case of a company looking at an industry barreling down one path and asking, “Wait, does this actually need to be so complex?” The entire trend has been towards more orchestration, more control frameworks, more deterministic workflows. Brex is betting that’s over-engineering for a future where the AI models are good enough to just talk to each other like people on a team. The risk, of course, is chaos. How do you debug a system with no central log? How do you ensure consistency? Brex says they’ve baked in evaluation and audit agents, and the entire MessageStream is logged. But it’s a fundamentally different way to think about building software. If they’re right, it could make a lot of current AI architecture look clunky. If they’re wrong, well, it’s a fascinating experiment in distributed AI. Either way, the ambition to become invisible is probably the right one for enterprise tools. The best technology doesn’t feel like technology at all.
