Agency.
The company building the tools that let you see what your AI agents are actually doing.
// The logo of a startup whose first product was a debugging tool that won the hackathon it was built for, then refused to stop growing.
The company building the tools that let you see what your AI agents are actually doing.
// The logo of a startup whose first product was a debugging tool that won the hackathon it was built for, then refused to stop growing.
It is 2 a.m. somewhere in San Francisco, and an AI agent is writing a company's blog post without supervision. Somewhere else, an agent is reconciling a hedge fund's data. Most of the time these agents work. Sometimes they do something strange, and nobody can say exactly what or why. Agency exists for that gap - the distance between "the agent ran" and "we know what it did."
Agency is a small company with a deliberately oversized job: making autonomous software legible. Its flagship product, AgentOps.ai, sits quietly underneath AI agents and records everything - every prompt, every tool call, every decision, every dead end. The pitch is almost boring in its sensibility. While much of the industry races to build flashier agents, Agency decided to build the instruments that tell you whether your agent is any good. The unglamorous middle of the stack, it turns out, is where the money and the anxiety both live.
"Instead of building an agent ourselves, we should build tools to make it easier to build agents."
- Alex Reibman, Co-Founder & CEOHere is the inconvenient truth about AI agents: they are confident and frequently wrong. In the summer of 2023, Alex Reibman was building web-scraping agents that failed somewhere between 30 and 40 percent of the time. That is not a rounding error. That is a product that breaks during the demo. The problem was not that the agents failed - software fails - it was that when they failed, the failure happened inside a black box. No stack trace. No replay. Just a wrong answer and a shrug.
An agent is not a single function call. It is a chain of reasoning steps, tool invocations, and model outputs, each one capable of quietly poisoning the next. Debug that with print statements and you will age a decade. The teams shipping agents into production - hedge funds, consultancies, marketing firms, the occasional Fortune 500 - were flying instruments-down. They needed a flight recorder. There wasn't one worth using.
"Agents represent new members of the workforce requiring the same auditability."
- Adam Silverman, Co-Founder & COOSilverman's framing is the whole thesis in one sentence. In a large enterprise, you know what your employees do - there are logs, reviews, audit trails, someone to ask. Agents are new hires with no paperwork. They make decisions that cost money and carry liability, and until recently they did it with no record anyone could inspect. The compliance department, it should be said, finds this less charming than the engineers do.
The origin story is almost too neat. Reibman built a debugging tool to troubleshoot his misbehaving scrapers, entered it in one of the city's AI hackathons, and it won. Then something more telling happened: other builders asked if they could buy it. The tool he made to scratch his own itch turned out to be everyone's itch. So Reibman, alongside co-founders Adam Silverman, Shawn Qiu, and Braelyn Boynton, made the contrarian call. They would not build another agent. They would build the layer that watches all the others.
It is a quietly clever position. Every new agent framework that launches - and they launch weekly - is a potential customer rather than a competitor. AgentOps is model-agnostic and framework-agnostic by design, which is the polite way of saying it refuses to bet on which agent platform wins. It just wants to be present when any of them runs. The team calls Agency a consulting firm and a product company at once: agen.cy has built and reviewed hundreds of custom agents for startups and large enterprises, which conveniently doubles as the world's best market research for what the product should do next.
// A company aged in hackathon-years, where six months counts as an era.
AgentOps does one thing with unusual seriousness: it logs every interaction and decision an agent makes, then makes that record useful. Developers can watch sessions replay step by step, see exactly where reasoning went sideways, measure cost and latency per run, and set guardrails that stop an agent before it does something expensive. Silverman calls it "multi-device management for agents" - a fleet-management console for software employees who never sleep and occasionally hallucinate.
The integration story is the part developers actually care about. AgentOps plugs into Microsoft's AutoGen, CrewAI, LlamaIndex, Cohere, Mistral, and MultiOn - usually in a couple of lines of code. The SDK lives in the open on GitHub, which is how a tiny team reaches thousands of monthly teams: let the developers find you, then earn the enterprise contract when their prototype becomes a production system with a compliance officer attached.
// Where the attention went vs. where the failures hide. Bars are illustrative of Agency's thesis, not audited metrics.
// The gap between "everyone builds agents" and "almost nobody can audit them" is the whole business.
In August 2024, Agency raised $2.6 million in pre-seed funding led by 645 Ventures and Afore Capital. The investors did not bury the lede. One described the company as "Stripe for payments, or Twilio for communications" - the unsexy primitive that an entire category quietly depends on. That is the dream multiple: become so foundational that builders reach for you without thinking, and switching away feels like ripping out the plumbing.
"AgentOps is cracking that code, drastically speeding up development time - think Stripe for payments, or Twilio for communications."
- Afore CapitalThe customers are the more convincing evidence. Thousands of teams a month run their agents through AgentOps - hedge funds tracking trading research, consultants auditing client deployments, marketing firms keeping their content agents honest. The most quietly delightful one: a company that uses AgentOps to babysit an AI agent that writes its blog posts, presumably so a human finds out before the internet does when the robot goes off-message.
Agency's stated ambition is larger than dashboards. Reibman calls safe, accessible, and scalable AI agents "the ambitious project of this generation," and the company's whole posture follows from taking that seriously. If agents are going to do real work - move money, ship code, talk to customers - then someone has to make them auditable, debuggable, and accountable. Not because it is exciting, but because no serious enterprise deploys software it cannot inspect. Agency is betting the boring requirement becomes the gating requirement.
"Safe, accessible, and scalable AI agents will be the ambitious project of this generation."
- Alex Reibman, Co-Founder & CEOEvery month, more companies hand more decisions to agents. Each one widens the gap Agency was built to close - the gap between what software does and what we can prove it did. As agents move from writing blog posts to touching money and infrastructure, "we think it worked" stops being an acceptable answer. The observability layer stops being a nice-to-have and becomes the thing the auditor, the regulator, and the on-call engineer all demand first.
Return to that 2 a.m. scene. The agent is still writing the blog post, still reconciling the data, still running while everyone sleeps. The difference now is that there is a record. When morning comes, someone can open AgentOps and read exactly what the machine decided, where it hesitated, and what it nearly got wrong. Agency did not make the agents smarter. It made them legible. In a world filling up with software that acts on its own, being able to read it back may be the more valuable trick.