Hiring an 'Agent Engineer': JD and Skills Profile

The role exists, sort of. The skills, the interview signals, and the JD template, with the parts that should differ between platform teams and product teams.

The skills profile

Four skills define the agent engineer. Strong software engineering (the agent is a software system, not just a prompt; engineering rigour matters); evals fluency (knows how to design test suites, score outputs, detect regressions; this is the rare skill); prompt engineering (writes and refines prompts; less rare than evals but still uncommon at depth); production operations (knows what observability, on-call, and SLOs mean, otherwise the agent ships without them).

• Strong software engineering. Agent is a software system; engineering rigour matters.
• Evals fluency. Test suites, scoring, regression detection; the rare skill.
• Prompt engineering. Writes and refines; less rare but still uncommon at depth.
• Production operations. Observability, on-call, SLOs; otherwise agents ship without them.

Interview signals

Four signals separate real candidates from pretenders. “Tell me about an eval suite you built” (real candidates have stories, pretenders have generalities); “how do you debug a prompt that started failing in production?” (real candidates describe a methodology, pretenders describe symptoms); “what is the cost trade-off between Haiku and Opus for triage?” (real candidates have an opinion with reasons, pretenders quote vendor pages); take-home design an agent for X workflow.

• Eval suite story. Real candidates have specific stories; pretenders speak in generalities.
• Prompt-failure debugging methodology. Real candidates describe a process; pretenders describe symptoms.
• Haiku vs Opus cost trade-off. Real candidates have a reasoned opinion; pretenders quote pages.
• Take-home design. Small coherent design beats long wishlist; the bake-off.

JD for platform team

The platform-team JD is infra-flavoured. Focus: build the agent platform that other teams use; think infrastructure engineer with LLM specialty. Skills: distributed systems, observability, API design with LLMs as one component among many. Compensation: senior software engineer level with LLM premium of 10-15%.

• Build the agent platform. Other teams use it; the platform team is the foundation.
• Infrastructure engineer with LLM specialty. The core archetype; LLM is one component.
• Distributed systems plus observability plus API design. The core skills; LLM is layered on.
• Senior plus 10-15% LLM premium. The compensation anchor.

JD for product team

The product-team JD is workflow-flavoured. Focus: build agents for specific workflows; think feature engineer with LLM specialty. Skills: domain expertise, prompt engineering, eval design with LLMs as the core tool. Compensation: senior software engineer level with domain premium based on the product area.

• Build for specific workflows. Feature engineer with LLM specialty.
• Domain expertise required. Product area knowledge plus prompt engineering plus eval design.
• LLMs as core tool. Different from platform team; the LLM is the primary instrument.
• Senior plus domain premium. Compensation reflects domain area.

Retaining them

Three things keep agent engineers engaged. Give them ownership of the eval suite (without owning quality, they cannot improve it); give them visibility into the impact (their work changes MTTR, show the chart); give them time for craft (prompt and eval engineering reward iteration, under-resourcing produces mediocrity).

• Eval suite ownership. Without owning quality, they cannot improve it.
• Impact visibility. Their work changes MTTR; show the chart.
• Time for craft. Prompt and eval engineering reward iteration; under-resourcing produces mediocrity.
• Per-engineer growth path. The role grows with the agent platform; supports retention.