About the role

We already run two kinds of agent alongside our teams. AI agents diagnose performance problems and tell people the root cause. Rule-based agents recommend marketing actions to our advertising teams. The quality of their decisions shows up directly in our revenue.

Next, both take a greater role in the day-to-day running of our lead generation business, moving from recommending, to acting with a person's approval, to acting on their own where it is safe. We need an engineer to build agents that are dependable in production, measurable, and safe to trust with more.

This is a hands-on engineering role. You'll own agents end to end, from understanding the workflow to running them in production, alongside our existing AI engineer, data scientists, product managers and the people who use the agents every day. Today we use Claude and open-weight models such as GLM, chosen per task on quality and cost, on an agent framework we build ourselves.

What you'll do

Build agents

Start from the workflow. Sit with the operators, map how the work is actually done, and build for that, not for the process as documented.
Build on shared rules. Rule-based and LLM agents that call shared, versioned rules and tools, not logic hard-coded inside each agent.
Connect them to the business. Using our data warehouse, internal services, and the APIs into the systems agents act in.
Turn outputs into results people trust. Structured outputs, validation and clear application logic.

Make them safe to trust

Log every decision. Each action records its context, what was chosen, and whether a person accepted or overrode it, so agents can be evaluated and improved.
Build the path to autonomy. Per decision, agents move from recommend, to act with approval, to act alone. Our data scientists propose the bar from agreement and override data and point people in the business sign it off; you build the mechanism that enforces it on our shared agent framework, with permissions and guardrails in the code, not the prompt.
Evaluate behaviour, not just code. Test sets, agreement with expert operators, and evaluation before and after every change.
Run what you ship. Tracing, failure handling, and tracking quality, latency and cost. There's no on-call rota.

What we're offering you:

Summer Fridays
Competitive holiday benefits - 25 days a year paid holiday, plus 8 bank holidays (increases 1 day a year up to 30 days)
Hybrid working - 3 days a week in the office
Closed for Christmas holidays - Extra days not taken from your annual holiday allowance.
Work from anywhere for 2 weeks a year
Life Assurance and Income Protection to protect your loved ones
Benefits allowance for health, dental, and vision coverage
Six months paid maternity leave, and one month paid paternity leave (subject to qualifying conditions) inclusive of same-sex and adoptive parents
Defined Contribution Pension and Salary Sacrifice Scheme
Be Well: Our award-winning wellbeing and mental health programme to support all MVFers and their families
Family Forward support for our MVF parents and their mini-mes
2 charity days a year

What we're looking for

3+ years of engineering, including at least one LLM application or agent shipped to real users.
Python and TypeScript: production-grade in one and productive in the other, with typed, tested code and clear interfaces. You'll work in both here.
LLM engineering: tool calling, structured outputs, context management, and evaluating agent behaviour on representative test sets.
Integration and operations: SQL, HTTP APIs, containers, CI/CD and a cloud platform, and debugging from logs and traces.

Nice to have

Agents used daily by operators, with a feedback loop.
AWS SageMaker, AWS ECS/Fargate, Terraform and CircleCI.
Snowflake and dbt.
Event logging for analytics or machine learning.
Lead generation or ad tech.

How you'll work

Commercial impact. You understand how the business creates value and use that to make better choices about what gets built, prioritised or stopped. You define the intended impact before starting, consider the cost of the work and evaluate what happened afterwards.
Delivery and ownership. You take responsibility for outcomes and move at a pace matched to the cost of delay. You follow work through until it is delivered, handed over or deliberately stopped, and remain accountable for what happens after delivery.
Candour and communication. You surface important information early, including when it is uncomfortable, with evidence and a view on what to do about it. You communicate what is known, assumed and uncertain in a form the audience can understand and act on.
Making others better. You create lasting improvement in other people's and the team's work through teaching, feedback or challenge, and take challenge on your own work as readily as you give it.
Growth and mastery. You deepen your capability through deliberate practice, feedback and experience, build the skills the work will need next, and change established practices when better approaches become available.
Judgement under ambiguity. You make sound decisions with incomplete information and know when to act, escalate or delegate. You own what you ship regardless of how it was produced, and change position when the evidence changes.

How you'll build

Craft and AI-native delivery. You build robust agents and AI applications, choosing the right approach for the problem, including when deterministic code is enough. You understand and can explain the evidence, assumptions, behaviour and limitations of what you build, and use AI to increase delivery speed without giving up technical ownership.
Architecture and systems design. You design reliable systems that move agents and AI capabilities from a business problem into effective use, deciding what should be probabilistic and what deterministic. You treat data and context, security and governance, monitoring and observability, failure handling and recovery, cost, performance and lifecycle as part of the design rather than concerns added afterwards.

AI Engineer