How to hire an AI agent developer: questions to ask, red flags and what you should own.

01What does an AI agent developer do?

An AI agent developer designs, builds and maintains software agents that use a large language model to carry out multi-step work with your tools and data. The job covers prompts and code, plus the tests, limits, monitoring and documentation that let a business trust an agent near customers and money.

An AI agent developer is an engineer who builds AI agents: software that uses a large language model to plan and complete multi-step tasks with a business's tools and data, inside limits the business sets, with tests and monitoring that show it works.

Typical marketing agents research accounts, draft first messages, audit a website for AI answers, or write the weekly report. The glossary entry on agentic AI covers the wider term, and the guide to AI marketing agents shows how agents fit into one system.

You will be competing with employers for these people. A SEEK New Zealand search for AI agent developer jobs in October 2026 returned AI engineer, AI developer and contract AI automation roles¹ across the country.

02AI agent developer, AI engineer or AI developer: what is the difference?

The titles overlap, and job ads use them loosely. AI developer is the broadest label, for anyone writing software that uses AI. An AI engineer usually builds and deploys models or AI features inside a product. An AI agent developer specialises in agents that take actions in business tools, so the work leans toward integrations, permissions and review steps.

Salary guides use the engineer titles. Robert Walters' 2026 salary guide for Australia and New Zealand lists AI engineer and AI consultant roles for Auckland, and separate generative AI engineer and machine learning engineer roles for New South Wales². None of its tables uses the title AI agent developer, so hire against the work you need done.

03Freelancer, agency, in-house hire or contract engineer: which fits?

A freelancer suits one well-defined agent. An AI agent development company suits several agents across teams, with more people to call when something breaks. An in-house hire suits a business that will build agents for years. A contract engineer suits a fixed build with a handover, usually billed by the hour through a recruiter.

Robert Walters' 2026 guide puts an Auckland AI engineer at NZD 130,000 to 150,000 base salary, or NZD 130 to 180 an hour on contract, and a New South Wales AI engineer at AUD 150,000 to 170,000 or more before superannuation. SEEK's advertised average for a New Zealand software engineer in information and communication technology was NZD 115,265 in October 2026, across 648 openings³. In New Zealand, add employer KiwiSaver, which Inland Revenue confirms rose from 3% to 3.5% on 1 April 2026⁴.

Freelancers and agencies that build AI agents rarely publish rates, and this guide found no survey of them to verify, so ask for a written quote per agent.

The AI marketing cost guide for Australia and New Zealand sets these figures against retainers and tools, and agency vs freelancer vs in-house covers the trade-offs.

04Which skills should an AI agent developer show you?

Look for five skills beyond prompt writing: evaluation, guardrails, data access design, monitoring and documentation. Each one leaves an artefact you can inspect, such as a test set, a permissions list, a run log or a runbook. A developer who has these skills can open examples during the first call.

Evaluation

An agent needs a test set: real inputs with the outputs a good result should contain, run before every change. MBIE's voluntary Responsible AI Guidance for Businesses, published July 2025, says AI systems should be tested often and actively monitored, and advises setting success metrics and thresholds to evaluate against⁵.

Guardrails

Guardrails decide what the agent may do alone. Drafting is usually fine; sending, publishing, spending and deleting should wait for a person, known as human in the loop. Agents that read emails or web pages take in text written by strangers, and the OWASP Gen AI Security Project ranks prompt injection first in its 2025 Top 10 for large language model applications⁶. Ask how the developer stops a hostile email from steering the agent.

Data access design

Good developers give each agent the least access that does the job: read-only first, named fields rather than whole databases, a separate service account where possible. Ask for the list of systems each agent can read and write.

Monitoring

An agent that worked in testing can drift once it meets live data. Australia's National AI Centre, in its Guidance for AI adoption, tells organisations to monitor AI systems after deployment to detect changes in performance and behaviour⁷. In practice: a log of every run, an alert on failure, and a weekly sample a person reads.

Documentation

Each agent should have a one-page brief (its job, inputs, tools, limits and reviewer), a runbook for when it breaks, a list of the accounts and keys it uses, and a change log.

05What should you ask before you hire?

Ask questions that have the candidate show working examples. Eight questions cover the ground: a live run, the test set, the limits, the data, failure handling, monitoring, ownership and handover. Put the same questions to every candidate, in writing where you can, and score the answers on the table further down.

1. Can you show me an agent you built, running on real work? Input, steps, output and review, live or recorded.
2. What is in its test set, and when do you run it? Good answers name the awkward cases and run tests before every change.
3. What can the agent do without a person? Look for a written list, with anything customer-facing behind a review.
4. Which of our systems will it read and write, and where is the data stored? Ask for the country and provider of each tool, including the model.
5. What happens when it fails? Expect alerts, a safe fallback and a person who gets told.
6. How will we know it still works in three months? Logs, a metric with a threshold, a regular sample review.
7. Who owns the code, prompts, test sets and documentation? Your business, in writing.
8. What does the handover include? A list, a walkthrough and a defined support period.

Australia's National AI Centre agrees: if you are buying an AI system, its guidance says to ask the developer or supplier to show proof that it has been properly tested⁷. The guide on how to choose an AI marketing provider applies the same test to agencies.

06What are the red flags?

The clearest red flags are a developer with no running agent to show, no test set, agents built on the developer's own accounts or API keys, requests for broad access to save time, promises that an agent needs no human review, and vague answers about where your customer data goes. Two of these usually justify looking elsewhere.

• A slide deck or a tool demo in place of an agent they built
• The model account, automation platform or code repository in the developer's name
• An agent that sends, publishes or spends with nobody reviewing
• A contract silent on IP, or one that licenses the work back to you

Gartner warned in June 2025 about "agent washing", estimating that only about 130 of the thousands of agentic AI vendors are real, and predicted over 40% of agentic AI projects will be cancelled by the end of 2027⁸. A running agent with tests settles the question faster than any title.

07How do you score candidates side by side?

Score each candidate from 0 to 3 on six criteria, multiply by the weight, and add the results. The weights favour what is hardest to fix after you sign: a working agent, evaluation, limits on access and clear ownership. Compare price only between candidates who pass those first four.

The maximum is 45. A candidate under 27 needs a strong reason to stay on the list.

08What must the contract say about IP, accounts and data?

The contract should assign you the code, prompts, test sets and documentation, require every account, API key and connector to sit in your name, set which data the agents may read and where it is stored, and describe the handover. Default copyright rules differ between New Zealand and Australia, so write ownership down either way.

Intellectual property

In New Zealand, section 21 of the Copyright Act 1994 makes a person who commissions and pays for a computer program its first owner, subject to any agreement to the contrary⁹. That list names computer programs but not written works in general, so assign prompts, instruction files and documentation to your business in the contract too, and treat this as a practical reading of the Act to confirm with a lawyer. In Australia the default runs the other way: IP Australia says IP created by a contractor is the property of the contractor unless otherwise stated in the contract¹⁰. If the developer reuses their own libraries, get a permanent licence to anything your agents depend on.

Accounts and keys

The AI model account, automation platform, hosting and code repository should belong to your business and bill your card, with the developer added as a user and API keys created in your accounts. When the engagement ends, you remove their access and the agents keep running.

Data handling in New Zealand

The Privacy Act 2020¹¹ applies to any business handling personal information, and agents that read a CRM or an inbox handle plenty. The Privacy Commissioner's information privacy principles cover storage and security (IPP 5), limits on retention (IPP 9) and disclosure outside New Zealand (IPP 12)¹², and the Commissioner expects a privacy impact assessment before AI tools are used, updated regularly¹³. Agents that enrich lead lists also fall under IPP3A, in force since 1 May 2026, on telling people about indirect collection¹⁴.

Data handling in Australia

The OAIC's summary of the Australian Privacy Principles says APP 8 sets the steps before personal information is disclosed overseas, and APP 11 requires reasonable steps to protect it from misuse, loss and unauthorised access¹⁵. The OAIC defines a small business as one with annual turnover of A$3 million or less¹⁶, and most of those sit outside the Act, though health providers and some others are covered regardless. The privacy law guide for NZ and Australia goes deeper.

09What does a good handover and exit look like?

A good handover leaves your team able to run, change and switch off every agent without calling the developer. That means the repository in your organisation, a runbook per agent, the test set, a list of accounts and keys, a recorded walkthrough, and a defined period of support. Agree all of it in the scope, before work starts.

1. Code, prompts, instruction files and test sets in a repository your business owns
2. A runbook per agent: how to pause it, who to call, how to roll back a change
3. A list of every account, key and connector, with the owner of each
4. A recorded walkthrough with the person who will run the agents
5. A support period with a response time, priced in the original quote
6. On exit: developer access removed, keys rotated, and the developer's copies of your data deleted

The last item matters under IPP 9 as well as for security. Record the list and the exit terms in a written AI marketing scope of work.

10Where Flow AI fits

Flow AI provides AI marketing services for businesses in New Zealand and Australia. Amber Lan, an AI marketing engineer in Auckland, builds the AI agents that run search, content, outbound and reporting with your team, then hands them over. The AI agent development service sets out how a build runs on your accounts, with a named person on your team approving each agent's work.

The free agent templates and the public skill library show the method before a call. Each build starts with a written scope and is priced per project. Companies hiring a marketing engineer or GTM engineer can reach Amber through the hiring page.

11The short version

Hire an AI agent developer on evidence: an agent running on real work, a test set, written limits on what it can touch, monitoring you can read, and documentation another developer could pick up. Score every candidate on the same table, and compare price only among those who pass the first four rows.

Then put ownership in writing: code, prompts, tests, accounts and data in your name, privacy duties and the handover agreed before the build begins.

To put Flow AI through the same questions, talk to Amber.