20 Years at Goldman Taught Me How to Manage People. Turns Out, Managing AI Agents Isn't That Different.

‍

TL;DR: Managing AI agents draws on the same principles as managing people: clear roles, feedback loops, earned autonomy, and institutional memory. After 20 years as a Managing Director at Goldman Sachs, I've found that my top people management lessons translate almost directly into how we build and run agent systems at Genesis. Here are my 10 insights, with one rule of thumb: wherever you see the word "agent," you can swap in "team member" and it still holds.

From the Trading Floor to Agentic Systems

I spent 20 years as a Managing Director at Goldman, so I suppose I should know a thing or two about "managing" and "directing."

Now I'm building an Agentic Data Engineering platform, and interestingly, many of the same lessons carry over.

It turns out that managing agents and humans is not that different. Here are my top 10 agent and people management insights. You can replace the word "agent" with "team member/employee" and it still holds.

The 10 Lessons

1. Define narrow roles, then connect them. The more you ask one agent to do everything, the more it becomes a flaky genius who burns out. The clearer you define narrow roles and glue them together with thin orchestration, the more the whole org/agent swarm actually ships.

2. Build in reflection pauses for long-running work. The longer the mission, the more a single unchecked mistake turns into a project failure. The more you require reflection pauses, peer review steps, and "ask your manager" escalation, the further your agents go without exploding. This is a core reason we structured Genesis Missions around staged checkpoints rather than open-ended runs. For a deeper look at how context compounds over long tasks, see our post on Context Management: The Hardest Problem in Long-Running Agents.

3. Focus beats context overload. Overwhelm your agent with context (think: many emails and endless Slack threads) and the less productive it becomes. Give it a focused task with just the right amount of context, let it run, and watch it deliver.

4. Visibility prevents end-of-project surprises. The less visibility you have into every decision your agent makes, the more likely you are to be surprised at the end result. The better your meeting notes (full traces and replays), the faster you have clarity on what happened. This is one reason replay functionality is built into the Genesis platform: every step an agent takes is reviewable after the fact.

5. Evaluate early, not just in production. The later you wait to measure if they're actually good at their job, the more you discover in production that they've been making up KPIs. The earlier you run evaluations on real work results, the fewer times you have to fire (retrain) the whole team. This connects directly to how we approach synthetic data generation as a testing tool, getting realistic results before anything touches production.

6. Lean orchestration beats big debates. The bigger you make the "let's have fifteen agents debate for six hours" stand-up, the higher the token bill and the less gets done. The more you replace debate with single-responsibility agents and deterministic workflow, the leaner and more profitable the department. Microsoft's architecture guidance makes a similar point: distributing work across specialized agents requires careful observability and resource tracking to avoid runaway costs.

7. Validate outputs, always. The more you let agents use tools without double-checking the output, the more they confidently ship garbage to customers. The stricter your output validation and "show your work" policy, the fewer 2 a.m. apologies you send.

8. Treat agents like ambitious juniors, not self-managing seniors. The more you treat agents like magical self-managing seniors who improve forever without feedback, the faster they quietly accumulate bad habits and technical debt. The more you manage them like ambitious but erratic juniors, constant feedback loops, evaluations, and iteration, the longer they stay useful. 

9. Earn autonomy before granting it. The more autonomy you grant before trust is earned, the more spectacular the disaster when they go rogue. True autonomy is the final promotion: start them on a tight leash, graduate through staged approvals and human-in-the-loop gates, and only cut the cord when the evaluations, traces, and track record prove they deserve it. See how we put this into practice with Progressive Tool Use and the Agent Server access control model.

10. Manage long-term memory deliberately. The less you manage their long-term memory, the more they turn into brilliant new-joiners who re-debug the same broken DAG and re-profile the same 10 TB table every single morning. The more you deliberately curate, prune, summarize, and inject a clean, versioned institutional memory, the closer they get to the mythical 20-year principal engineer who actually remembers why that join exploded in 2022 — and never makes the same mistake twice. Cloudflare's recent work on Agent Memory explores this challenge at the infrastructure level; it's one of the fastest-moving areas in the whole space.

The Philosophy Behind Genesis Data Agents

The best systems, human or artificial, aren't built on blind trust or micromanagement. They're built on clarity, feedback loops, and earned autonomy. That's the philosophy we apply when building Genesis Data Agents. To see what this looks like in practice at enterprise scale, the GXS Bank case study is a good place to start: a team that cut data pipeline development from months to hours by applying exactly these principles.

Frequently Asked Questions

Does this mean AI agents will eventually replace data engineering teams? No. The analogy actually points in the opposite direction: good management makes a team more effective, not redundant. Genesis agents handle the repetitive, high-volume work so engineers can focus on architecture, modeling, and strategic decisions. Think augmentation, not replacement. The Junior Data Engineer is Now an AI Agent post covers this shift in more detail.

What does "human-in-the-loop" actually look like day to day? It varies by task complexity and how much trust a given agent has earned. Early in a project, agents surface questions and wait for input before proceeding. Over time, as their track record builds, the number of checkpoints decreases. Genesis Missions are structured around this graduated model by design.

How do you handle institutional memory across agent sessions? This is one of the harder problems. Agents don't retain memory between sessions by default, which is why we invest heavily in context graphs, curated knowledge injection, and versioned documentation. The goal is the same one you'd have with a human team: making sure the organization doesn't have to relearn the same lessons every time someone new joins.

Where should I start if I want to see Genesis in action? The walkthrough series is the fastest on-ramp. Walkthrough #1 starts with exploring a data source, and the series walks through a full mission end to end.