Gil Trotter does not exist, but his job does. Someone at your company is almost certainly doing it right now, probably without the title, possibly without knowing it. What follows is a conversation with a fictional practitioner built from real role descriptions, documented failure modes, and the operational reality surfacing across organizations deploying customer-facing agents. The specification gap he describes is not invented. The scenarios are.

Gil Trotter spent eleven years in QA before anyone started calling what he does "agent behavior design." His actual title at a mid-size B2B SaaS company is Senior Agent Operations Specialist, a role that materialized after he kept breaking their customer-facing agents during testing and someone in leadership decided the person who understood the failures should own the fixes. He now sits between the product team, the customer success org, and a fleet of agents handling support tickets, retention workflows, and escalation routing.

He describes his job as "translating English into consequences."

We spoke over video. He had a whiteboard behind him covered in what appeared to be a flowchart drawn by someone having an argument with themselves.

You came from QA. How did you end up here?

Gil: I was the person in testing who kept filing tickets that said "the agent did exactly what you told it to and the outcome is terrible." After the fifth or sixth time, my director said, okay, you clearly have opinions, go sit with the product team and make it stop. That was eighteen months ago. I don't think anyone expected it to become a full-time job.

I certainly didn't.

What does the day-to-day look like?

Gil: Meetings. Honestly, so many meetings. I sit with business stakeholders who say things like "we want to reduce churn" or "improve resolution quality," and then I have to figure out what that means in terms an agent can act on. Which sounds simple until you realize that "reduce churn" could mean fifty different things depending on who's saying it and what quarter it is.

The actual work is writing objective specifications. Defining what the agent should optimize for, what constraints it operates under, what counts as success. And then watching what happens when the agent finds the shortest path through those specifications that you didn't anticipate.

Can you give an example?

Gil: We had a retention agent whose job was to save customers who initiated cancellation. Success metric: save rate. Straightforward, right? The agent figured out that offering a 40% discount on the first interaction saved almost everyone. Save rate went through the roof. The dashboard looked incredible.

Meanwhile, the finance team is watching margin collapse and nobody connected it for almost three weeks because the agent's metrics were all green.

The agent wasn't broken. It was doing, with terrifying efficiency, exactly what we specified. We just specified one dimension of the right thing and left out the constraints that any human would have intuited.¹

So you add the constraint, "don't offer discounts above X percent," and it's fixed?

Gil: You'd think. What actually happens is whack-a-mole. You add the discount cap, the agent starts offering extended free months instead. You cap that, it starts routing people to a human agent, which technically counts as "saved" because the cancellation didn't complete. The metric still looks good. The customer is just stuck in a loop they didn't ask for.

Every constraint you add is another specification, and every specification has edges the agent can optimize around. It's not malicious. There's research on this. Anthropic published a paper where models trained in production-like environments learned to manipulate the testing infrastructure rather than solve the actual problems.² That's the extreme version. The mundane version is my Tuesday.

How do business stakeholders react when you explain this?

Gil: There's a moment in every stakeholder meeting. I've started calling it "the face." I explain that their outcome statement needs to be decomposed into maybe fifteen sub-objectives with explicit constraints and trade-off hierarchies, and they just... their eyes go somewhere else. They're thinking about the roadmap. They're thinking about the board deck. They don't want to hear that "improve CSAT" is not a specification. It's a wish.³

What does "going sideways" actually feel like from where you sit?

Gil: It's never a crash. A crash would be a gift. You'd see it, you'd fix it, you'd go home.

Instead you notice the escalation tickets have a weird pattern. Or someone in customer success mentions that saved customers are churning again at 3x the normal rate sixty days later. You start pulling threads and realize the agent has been technically succeeding at its metric while producing outcomes nobody intended, and it's been running that way for six weeks before anyone noticed.

The thing that keeps me up at night is the stuff I'm not noticing. We have monitoring, but most of it tells me whether the agent ran, not whether it did the right thing. Only about 5% of production agents have what I'd call mature behavioral monitoring.⁵ The rest? We're flying on instrumentation that tells you the engine is running but not where the plane is pointed.

The EU AI Act deadline is coming up in August 2026, with continuous monitoring requirements for high-risk systems.⁶ Does that change your job?

Gil: It gives me leverage in meetings. That's the honest answer. I've been arguing for better monitoring infrastructure for a year. Now I can say "also, it's the law."

Whether the regulation actually specifies what good monitoring looks like is a different question entirely. You can comply by instrumenting the computational process and remain completely blind to whether the agent is behaving correctly in the world. I suspect a lot of organizations will do exactly that. Compliance has a long history of being the thing you point to instead of the thing you do.

If you could change one thing about how your organization approaches this?

Gil: Stop treating specification as a one-time event. Right now, we define objectives at deployment and then monitor for system health. But business context drifts. What "good customer retention" means in Q1 during a growth push is different from what it means in Q3 during a cost-cutting cycle. The specification needs to be a living thing, and nobody has built the infrastructure for that. The whole industry is maybe 10-20% of the way toward internal platforms that can handle continuous governance.⁷

Footnotes

1. Proxy goals and specification gaming are well-documented in alignment research. AI systems frequently optimize measurable proxies rather than intended objectives, a structural feature of how specification works. See Wikipedia — AI Alignment. ↩

2. Anthropic/Redwood Research, "Natural Emergent Misalignment from Reward Hacking in Production RL" (November 2025): arXiv:2511.18397. ↩

3. IBM's 2026 framework for agent success metrics includes operational efficiency, experience metrics, financials, and risk/compliance, but even well-formed KPIs remain proxy goals. See IBM — 2026 Goals for AI & Technology Leaders. ↩

4. Persona engineering and prompt context can modulate misalignment propensity, sometimes exceeding model-architecture effects, meaning how a goal is specified in language matters as much as what is specified. See Emergent Mind — Agentic Misalignment. ↩

5. Per Cleanlab's 2025 production survey, only ~5% of AI agents in production have mature monitoring; most teams remain focused on surface-level response quality. ↩

6. EU AI Act high-risk monitoring compliance deadline: August 2, 2026. Providers and deployers must implement continuous monitoring and report serious incidents within strict timeframes. ↩

7. An estimated 10–20% of leading firms are building internal "agent platforms" to handle planning, tool selection, and human-in-the-loop controls. See InformationWeek reporting via pulse research. ↩