This isn’t a traditional senior engineering role. You won’t spend most of your time implementing product features directly.
Your time will roughly split:
50% building and evolving agent harnesses: orchestration, toolchains, approvals, secure execution, managed agents
50% reviewing and improving outputs: tracing failures, improving prompts/steering, tightening eval harnesses, reducing loop count
Concretely, you’ll:
Design and implement agentic workflows that take a requirement from spec → code → review → deploy
Build agentic loops that turn mistakes into system-level improvements (not one-off fixes)
Develop evaluation harnesses (offline + CI) to detect regressions in behavior, not just tests in code
Define and maintain review gates (human-in-the-loop + automated reviewers) for risky changes
Improve tool reliability: schemas, typed tool interfaces, retries, timeouts, safety checks
Build platform capabilities for managed agents: long-running sessions, checkpoints, state/memory boundaries, and recovery
Evolve the platform architecture (TypeScript, serverless architecture, shared codebase) with an eye for simplicity and maintainability
Partner with Product to reduce ambiguity and translate intent into testable, evaluable spec
This role requires strength in two areas, equally:
Systems thinking for agent harnesses and loops. You can design the execution harness around agents: feedback loops, evaluation strategy, safety constraints, and the “glue code” that makes autonomy safe in production.
Engineering taste. You can look at agent-generated code and immediately judge: conventions, simplicity, correctness, maintainability, security. Not just “does it work,” but “would I approve this PR in a regulated product?” What we need from you
Strong TypeScript and React experience in production environments
You’ve shipped real software to real users (not just prototypes)
You can read a codebase and quickly identify its patterns, conventions, and architecture
You are comfortable working in ambiguity and turning fuzzy intent into clear acceptance criteria + evals
Familiarity with agent tooling concepts: tool calling, MCP/tool integration, guardrails, evals, tracing/observability, and permissioning
Nice to have: AWS serverless experience (CDK, Lambda, DynamoDB). Our backend is a mix of modern serverless microservices and a legacy Express/PostgreSQL monolith. Who this role is not for
Be honest with yourself:
If you want to spend most of your time building features directly, this role will frustrate you
If you’re excited about AI but haven’t shipped production software, you won’t have the taste to judge agent output
If you prefer stable scope, established best practices, and minimal ambiguity, this environment won’t be a match The team and company
You’ll join a small team (3–4 engineers) reporting to a hands-on CTO. The company is going all-in on this model, not just engineering — sales, marketing, and support are all building agentic workflows for their functions.
This isn’t a side experiment; it’s our operating model.
We’re guided by trust, respect, and ownership. Our values, Embrace Change, Carte Blanche, Find Wisdom in Data, and We All “Own It”, shape how we work.
Fully remote (work from anywhere in Australia)
5 weeks annual leave and flexible working
Monthly Wellness Budget (mental & physical health)
Employee share options (ESOP) for all team members How to apply?
Submit your CV via the application form. Note that background checks are required as part of our offer process.
We welcome applications from all backgrounds, abilities, and identities. We value diversity and believe that it enhances our creativity, innovation, and overall success. Join us in creating a workplace where everyone can thrive.
Managing business expenses shouldn’t be a guessing game. Yet many SMEs still lack clear cash flow visibility and spend control. At Budgetly, we’re changing that.
We’re building an AI-first platform that simplifies expense management and helps businesses make better financial decisions, faster. The goal is simple: help you spend smarter, save time, and grow profitably.
Why this role exists (and why it’s different)
Six months ago, “agentic engineering” felt novel. Today, it’s rapidly becoming the default way teams ship software.
Most companies are using AI to make engineers type faster. We’re building something more ambitious:
We’re building an AI-first delivery system where agents ship product features end-to-end — and engineers build the platform (the harness) that makes it safe, reliable, and scalable.
If this works, feature delivery scales with product ambition, not headcount.
What we mean by “harness engineering”
A raw LLM isn’t an agent. The “agent” is the model plus the harness around it.
In this role, you’ll build that harness:
Agent runtimes & execution loops (plan → act → observe → reflect; retries; stop conditions)
Agentic loops & feedback loops that convert outcomes into improvements (evals, regressions, learnings)
Tooling & skills (MCP/tool integration, internal APIs, secure credentials, sandboxes)
Governance (permissions, policy, human-in-the-loop gates, audit trails)
Observability (traces, cost attribution, failure taxonomies, runbooks)
Evaluation harnesses (scenario suites, trajectory scoring, tool-arg correctness, “non-deterministic unit tests”) What success looks like in ~12 months
A managed agent platform that product can rely on for meaningful, customer-facing delivery
Agent workflows ship features with high repeatability, not “poke-and-hope”
Clear quality gates: eval harnesses, review agents, regression suites, and rollout controls
Engineers spend less time in the loop and on reviews and more on improving the factory (reliability, speed, safety)
