Staff AI Agent Infrastructure Engineer

Float is building an agent that works as a chief financial officer for every Canadian business: it knows the business, has all the context, and works around the clock. You would join the small team building the foundation for it, with a blank slate and a direct line to the product teams that will build on what you make.

You’ll have the best seat in the house on how to apply frontier AI to real finance problems: taking raw model intelligence and turning it into durable, scalable, useful products. The team sits between the model providers, Float’s product teams, and the customers whose finances we run, building the harness — memory, skills, tools, evaluations, observability, distribution channels — that every agent at Float runs on.

This is a zero-to-one position. You will build the core pieces that let agents act reliably in production, and the interfaces and channels customers reach them through, then scale them across Float’s products. The field is moving quickly, so the best people here learn fast and are glad to rebuild something when a better approach arrives.

We are hiring across multiple levels for this role. Our interview process evaluates scope, judgment, and learning velocity, and we level accordingly. At Staff, you set the patterns the harness is built on and the evaluation and observability bar every agent at Float meets. This role suits engineers who have shipped agents to real users, want more ownership than a large company can offer, and like to build.

What You'll Be Responsible For

• Ship agents that save customers real money and time. Take them from the first prototype to production at scale, and own the outcome once they’re live.

• Build the unified harness. Context management, capabilities and tool use, evaluation, monitoring, and observability — the layer every agent at Float runs on and scales through.

• Make agent actions safe to trust with a company’s money. Permissions, approvals, audit trails, and clear behavior when the agent is uncertain or wrong.

• Define what good looks like and measure it. Offline evaluations, production monitoring of agent decisions, and the loop that turns incidents into improvements.

• Build the interfaces agents are delivered through. Chat, generative UI, and multiple channels at once.

• Make product teams faster. Partner with them so they can launch their own agent capabilities on the harness without rebuilding it.

• Stay on the frontier. Track models, inference, and context management, and rework what we’ve built when a better approach appears.

• Representative projects:

  • Build the approval and audit layer that lets an agent pay a bill on a customer’s behalf.
  • Stand up the evaluation framework that tells us, before and after launch, whether an agent’s answers about a company’s spend are correct.
  • Give agents the context of a whole business — cards, bills, accounts, accounting integrations — in a form that stays accurate and token-efficient.
  • Ship the tooling that enables a product team to launch agentic capabilities without our help.