Want to know if this job is worth applying to?
Алматы, Қазақстан
Onsite
Almaty, Kazakhstan · Trust & Safety Engineering
Why this role exists
An agent needs more than a model response to complete useful work. It needs tools, durable execution, permissions, context, and a way to recover when something fails. You'll build the backend that lets Higgsfield agents carry creative tasks from a user's request to a finished result. We're looking for an engineer we can trust with a problem: work out what is worth building, build it, and carry it through to users.
What you'll actually do
Agent execution: build orchestration for multi-step tasks, tool calls, streaming updates, and cancellation.
Durable workflows: implement checkpoints, resumability, retries, and recovery without duplicating actions.
Tools and integrations: connect agents to generation, editing, and external services through clear, validated interfaces.
Execution boundaries: enforce user permissions, tenant isolation, resource limits, and approval requirements for consequential actions.
Evaluation and observability: trace failures, build regression cases, and improve task completion, latency, and cost.
Ownership: understand a real user task → define success and build evaluation cases → scope the simplest useful workflow → ship → measure completion, reliability, and dollar impact → improve from failed runs. Our release cycle is days, not quarters, so your first shipped agent capability happens fast.
Must-haves
You've taken a difficult problem from an unclear starting point to a working system used in production, and can explain your own decisions, contribution, and results.
You can build reliable agent execution: state machines, concurrency, queues, recovery, model APIs, secure tool calls, and repeatable evaluations. You learn unfamiliar tools quickly and understand what you ship.
You choose the next steps, remove blockers, and ask for help when needed. You cut scope to test an idea, adapt from evidence, and keep working through difficult problems.
You investigate real user tasks and failed agent runs, use feedback to choose what to automate, and stop workflows that do not help users finish their work.
You take feedback, change your mind when the evidence changes, and do the work the team needs. You share knowledge and responsibility so others can operate and extend what you build.
Who this is not for
You prefer deterministic application behavior without investigating model uncertainty.
You prefer prototyping without owning evaluation, recovery, and production reliability.
Nice to have
Production agents, workflow engines, or developer tools.
High-throughput systems with heavy traffic, large data volumes, or many long-running jobs.
Agent context management, memory, subagents, or parallel tool calls.
Sandboxed execution and scoped authorization.
Multimodal workflows involving images, video, or audio.
Interview process
Recruiter screen: role expectations, logistics, and fit.
Technical interview: a difficult system you shipped, what you personally decided and built, and how your work made the team more capable without leaving it dependent on you.
Additional technical interview, if needed: a deeper discussion of technical skills and open questions.
Product fit interview: scope a useful tool-using workflow, evaluate failures, and explain when user feedback would lead you to change or stop it.
Offer.
What we offer
Competitive base in USD, equity, relocation support, and ownership of systems that turn AI capabilities into completed creative work.
.png?1728912815){kind=logo}
.jpg?1700169058){kind=logo}