TRM Has a New Kind of Team Member. Meet Some of Our Roster.
Our legal team has an agent named Ruby the Redliner that performs first-pass contract redlining. Revenue operations built The Oracle, an AI forecasting engine that scores 2,000+ open opportunities in real time. Sales enablement runs a two-agent team, comprised of agents Noor and Vera, that produces weekly competitive intelligence briefs. Data partnerships has Mycroft, who generates integration code for new data sources.
Every one of these agents has a name, a defined role, a human partner, and clear boundaries around what it can and cannot do. And they're running in production today. None of them shipped overnight. Each one went through weeks — in some cases months — of iteration, failed experiments, and recalibration before earning the trust of the human it works alongside.
This post is about the operating philosophy behind how TRM thinks about humans and agents working together. It's also the opening chapter of a series where we'll introduce these agents one by one, told by the TRMers who built them. (For the broader context on TRM's AI transformation, read more here.)
The thesis: Agents as teammates, not tools
At TRM, we think of agents as teammates, not tools. That's a distinction that matters.
When you treat AI as a tool, you optimize for individual tasks: summarize this document, draft this email, clean this dataset. Useful, but limited. When you treat AI as a teammate, you design for something different: shared workflows, defined responsibilities, coordination across agents, and human judgment applied where it matters most.
At TRM, we've landed on the model “human-in-the-loop by design.” Every agent we've built has an explicit checkpoint where a human reviews, edits, and approves before the agent acts on anything externally-facing or high-stakes. For example, Ruby the Redliner doesn't send a redlined contract back to a counterparty without a lawyer's review. Noor and Vera produce weekly competitive briefs that go through a product manager's spot-check before they reach a battlecard. And the Oracle's forecast scores get validated against actual deal outcomes by the RevOps team.
We build agents to do the research, the first drafts, the pattern recognition, and the coordination; then route the decision to the human who owns the outcome. But we didn't start there. Every agent at TRM went through a period where its outputs weren’t correct or ready — where the human partner was iterating over and over, catching errors and feeding those corrections back into the system. The efficiency gains you'll see below are the result of that process, not a shortcut around it.
The part people get wrong about human/agent teaming
The most common misconception I hear from candidates, peers at other companies, and the broader discourse is that AI agents are about replacing people. That framing misses what's actually happening.
Every agent at TRM was built by a person who recognized that highly manual, repetitive tasks were taking an outsized share of their time — time better spent on the judgment calls only a human can make. Wei Wang didn't build Ruby because she wanted fewer lawyers on the team. She built Ruby because she wanted her team's time focused on strategic, high-impact decisions that actually require a lawyer's experience — not the mechanical first pass of identifying standard redline positions across a 40-page contract.
The pattern is the same everywhere we look: agents absorb the detail-intensive, repetitive execution so humans can focus on the craft. The research so you can focus on the insight. The first draft so you can focus on the edit. The coordination so you can focus on the decision. The work these agents do isn't low-value — it's essential: contract redlining protects TRM, competitive intelligence shapes deals, forecasting drives resource allocation.
The point isn't that this work doesn't matter. It's that the mechanical parts of it can be handled reliably by an agent that's been trained, tested, and corrected until it earns that trust.
That's the operating model we're building at TRM: humans working at the top of their capability, with agents handling the rest.
What's coming next
This post is the first in a series we're calling "Welcome to TRM, AI bots!" Over the coming weeks, we'll introduce TRM's AI agents one by one — their names, their roles, how they were built, what surprised their creators, and what they still can't do. Each post is co-authored by the TRMer who built the agent, because we think the most credible story about human/agent teaming comes from the people living it.
If building AI agents that do real work — not demos, not proofs of concept, but production systems with named roles and measurable output — sounds like your kind of challenge, we're hiring.
