Agent Beans

Agent Beans gives small businesses and indie developers an always-on growth team for Meta ads, built so they can trust it with real money. The US alone has 29.8 million one-person businesses and 6.4 million small businesses with employees. Many of them advertise on Meta, but few can afford an agency or an in-house team. As a result, the daily work of running ads, like reading yesterday's results, catching wasted budget, and testing new creative every week, gets done only when the owner finds time. Here is what Agent Beans fix: it gives each job to its own AI specialist, each running on a Meta model, while plain code gates every account change. The Bean Crew includes a Growth Lead, an Analyst, a Media Buyer, a Creative Strategist, a Creative Producer, and a Reviewer. Every day, they read the ad account, write findings, and propose budget and pause changes. Every week, they plan tests, research the market, and produce new image and video ads from the owner's own product photos. A seventh member, the Operator, is plain code that changes the ad account only after the owner approves in the Team Room, the app where the owner talks to the crew. A code harness is what makes it safe to hand over real money. While most AI tools for ads give general advice or act on the account by themselves, Agent Beans enforces in code what a prompt can only ask for. Every number must match an evidence ledger computed in code. Every market claim must link to a page the search actually returned. Six house rules, including Meta's learning phase and a maximum budget step, check every change. Approvals bind to a content hash, so what runs is exactly what the owner approved. Spend caps are checked before every model call, and every call is stored so it can be inspected exactly. Recorded runs and benchmark data prove its reliability. The live demo replays two sandbox sessions where Muse models made live decisions against simulated auction dynamics. On an offline evaluation suite of 6 scenarios across 3 random seeds featuring planted edge cases (wrong numbers, ungrounded claims, campaign-name prompt injections, post-approval tampering, and premature budget hikes), Agent Beans passed 18/18 runs, compared to 12/18 for an OpenAI Agents SDK baseline, 9/18 for LangGraph, and 3/18 for a single agent. Everything in the repository was built during the hackathon. Code: https://github.com/Jieyi-Deng/agent-beans Live demo: https://pallas-ads.github.io/agent-beans/