Capital Gambit

Inspiration

Every finance team asks the same question at the start of a cycle: "how do we earn more?" But that's the wrong question, and it's the one most FP&A tools are built to answer. The better question, the one an experienced CFO actually asks, is "where should we move our resources right now to get the biggest payoff later?" That's a resource allocation problem, not a growth-forecasting problem, and it's inherently multi-dimensional: you can't answer it by looking at cash alone, or costs alone, or revenue alone. You need all three, weighed against each other, at the same time. That's what a real finance team does with multiple specialists in a room. We wanted to build that room.

What it does

Capital Gambit is a multi-agent system that analyzes a company's cash position, cost structure, and revenue trajectory concurrently, then synthesizes those findings into one concrete, actionable capital reallocation recommendation for the next financial cycle.

A Cash & Working Capital agent analyzes AR/AP timing and the cash conversion cycle to surface specific levers (e.g., extend a vendor's terms, accelerate a customer's collections). A Cost Structure agent finds named, dollar-quantified inefficiencies in recurring spend. A Revenue/Forecast agent projects the next cycle and flags where the current forecast is drifting from actuals. A Synthesis agent takes all three analyses and proposes exactly where freed-up capital should move for the highest expected payoff, with the reasoning shown, not hidden, and routes it to a human for final approval. Nothing executes automatically.

Users connect their own financial data, a spreadsheet/ledger export and a live bank feed (Plaid sandbox) in this build, and get a clear recommendation instead of a dashboard full of numbers they have to interpret themselves.

How we built it

We used AO (Agent Orchestrator) as our build environment from the first commit. Each agent was developed as its own AO worker in an isolated git worktree, so we could build and iterate on the Cash agent, Cost agent, Revenue agent, and Synthesis agent in parallel without merge conflicts or context bleed between them. [Add specifics: number of AO sessions, which coding agent/model you ran inside AO, any CI/PR flow you used.] We deliberately kept the Synthesis agent's access limited to the outputs of the other three agents, not their raw reasoning, to keep its recommendation genuinely a synthesis rather than a rubber stamp of whichever agent spoke loudest.

Challenges we ran into

The hardest problem wasn't any single agent. It was making the Synthesis agent's recommendation trustworthy rather than just plausible-sounding. An LLM asked "where should we allocate capital" will happily generate a confident-sounding answer even with weak underlying signal. We had to build explicit reasoning requirements into the Synthesis agent so it has to cite which specific finding from which agent justified its recommendation, so a human reviewer can actually verify the logic instead of taking a black-box output on faith. [Add a specific technical challenge you actually hit, e.g. data normalization across sources, agent coordination timing.]

Accomplishments that we're proud of

We built a system where the human-in-the-loop step isn't an afterthought bolted on for compliance. It's structurally central. The Synthesis agent's recommendation is designed to be legible and challengeable by a real accountant, not just accepted. We're also proud that this isn't a single generalist agent pretending to understand finance holistically. It's genuinely decomposed the way a real finance team is organized, which came directly from thinking hard about how the Office of the CFO actually works, not just wrapping a prompt around "give financial advice."

What we learned

Multi-agent independence is easy to claim and hard to actually engineer. It requires deliberately restricting what each agent can see, not just giving them different names and prompts. We also learned that the quality bar for a finance recommendation tool isn't "does it sound smart," it's "can a skeptical accountant trace exactly why it said what it said," which pushed us to prioritize explainability over cleverness throughout the build.

What's next for Capital Gambit Wire up the remaining data connections (expense/subscription feeds, CRM/pipeline exports) so all three analysis agents pull from live sources rather than partial mock data. Add a feedback loop where the Synthesis agent's past recommendations get scored against actual outcomes, so its allocation logic improves cycle over cycle. Expand beyond a single "next move" recommendation to a ranked set of alternative allocations with trade-offs shown, so the human reviewer isn't just approving/rejecting one option but genuinely choosing between real strategic paths.