Who you are hiring
Prompt engineers optimize a turn. Agentic developers design systems: tool permissions, state, retrieval, interrupts, and traces. If a candidate can only name model brands, they are not yet the hire for a production agent.
What “production evidence” looks like here
When we hire or when you diligence us, the artifacts are public: Nexora’s DAG approvals, Conductor’s no-auto-merge rule, Overtone’s grounded speaking loop, Audit Genie’s reject-unanchored-claims filter. Ask any vendor—including us—for one system they owned under real load and one incident they debugged with traces.
A hiring sequence that wastes fewer weeks
- Written production story. Ask for one agent they shipped: goal, tools, gates, evals, and a failure. No story, no onsite.
- Whiteboard the graph. Have them draw state, retries, and the human interrupt. Listen for idempotent tools and bounded fan-out—not model trivia.
- Eval and cost. How do they stop a prompt change from breaking last month’s cases? How do they see spend before a run?
- Reference the repo. For agencies, require client-owned code and a handoff plan. For employees, require they can extend someone else’s agent, not only greenfield demos.
Signals vs red flags
| Look for | Walk away if |
|---|---|
| A production agent with users, logs, and a named failure mode | Only a weekend chatbot or a slide titled “multi-agent” |
| HITL on send, merge, deploy, or HRIS write | “The model is aligned so we do not need approvals” |
| Regression evals they can rerun | They “just eyeball a few prompts” |
| Cost or latency budget discussed unprompted | They cannot estimate calls per task |
| Clear ownership of the repo after kickoff | The agent lives only in the vendor’s account |
Bottom line: Hire the person or shop that can explain a boring failure. Charm and model-name fluency are cheap.
Agency or employee?
Hire an employee when you will run many agents and already have platform support. Hire an agency when you need the first production system and cannot staff two specialists this quarter. The checklist above does not change—only the contract does.
Related work and reading
Frequently asked questions
What interview question works best?
“Walk through one production agent you owned: requirements, architecture, tools, evals, an incident, and handoff.” Weak answers stay on prompts. Strong answers mention traces, permissions, and what they would not automate.
Should we test live coding of LangGraph?
A small graph with an interrupt node is useful. A leetcode tree problem is not. You are hiring systems judgment.
Can VisionsCraft help us hire after a build?
We can pair with your future owner during handoff. We are not a recruiting firm. Use the scorecard on this page with any candidate.
Is this the same as the blog checklist?
The blog is the essay. This landing is the durable hire page—table, sequence, and FAQs—linked from services and compare.