Note: The job is a remote job and is open to candidates in USA. ZON builds AI that improves foundational commercial models through recursive, autonomously adaptive processes and advanced AI orchestration. The Staff AI Engineer will design, deploy, operate, and evolve production-grade AI agents, developer harnesses, multi-model orchestration systems, and self-improving AI. The role also involves building evaluation, reliability, safety, cost-management, and AI-native development systems.
Responsibilities
- Design multi-step, tool-using, and multi-agent workflows (planner/executor, hierarchical, collaborative, human-in-the-loop) with explicit state, durable execution, retries, and safe fallbacks
- Build production harnesses: tool dispatch, session and state management, context as a managed resource, memory hierarchies, compaction, and recovery on long-running work
- Design self-improving loops: production traces → eval suites → prompt/skill/tool updates → regression gates → rollout, with a human in the loop on high-impact actions
- Use Temporal as the durable-execution backbone for long-running agent and workflow systems—signals, activities, retries, timeouts, versioning, and recovery across failures
- Route across commercial and open-source models for quality, latency, cost, capability, and policy—not for novelty
- Work fluently across the current stack of agent frameworks, flow tools, gateways, and protocols, including LangGraph, LangFlow, and Temporal, plus related commercial and open-source systems
- Integrate tools and internal systems through function calling, MCP, A2A (and related), hooks and APIs, retrieval, and structured outputs
- Build retrieval and context paths agents can ground in and act on
- Make coding agents, skills, and harnesses part of how software is specified, implemented, reviewed, and shipped
- Treat prompts, skills, graphs, flows, and Temporal workflows as production software: versioning, review, canaries, kill switches, audit logs
- Stand up evaluation as a first-class system—golden datasets, trajectory evals, automated graders where they are valid, human review where they are not
- Instrument traces, token and cost accounting, tool-error rates, loop detection, and drift. Write down what happens when the system misbehaves
- Design guardrails that survive contact with real users: least-privilege tools, injection resistance, PII handling, and clear authority at the skill and tool layer
Skills
- • You have shipped software other people depend on, including at least one LLM- or agent-backed system past the demo stage
- • You can design an agent or harness as a system: tools, state, context, evaluation, durable execution, and what happens when a step fails
- • Hands-on Temporal experience building or operating durable workflows—activities, signals/queries, retries, timeouts, workflow versioning, and failure recovery—especially as the execution layer under agents or orchestration
- • Extensive, current familiarity with both commercial and open-source AI tools, standards, models, and frameworks—not a single-vendor habit. That includes hands- on depth with LangGraph and LangFlow, and working knowledge across the surrounding ecosystem (orchestration and flow tools, model APIs and gateways, coding agents, eval/observability, MCP or equivalent tool standards)
- • AI-supported / AI-driven development is how you build. You use coding agents, harnesses, skills, and eval loops as the main path from spec to shipped change. You can drop to hand-crafted code when the problem demands it; you do not default to it
- • You still own the result: you can read, review, debug, and take responsibility for generated and orchestrated code, graphs, flows, and Temporal workflows
- • You have a real eval habit. You have caught a silent regression, written a grader, or refused to ship on a good vibe
- • You can take an underspecified workflow and leave behind an architecture, a way to measure it, and a clear account of the tradeoffs
- • Self-improving loops fed by traces and evals
- • Multi-model routing and cost/quality/policy tradeoffs across frontier and open/distilled models
- • CI and review for prompts, skills, graphs, LangFlow/LangGraph artifacts, and Temporal workflow definitions
- • Retrieval systems agents actually depend on
- • Guardrails, permissions, and auditability for autonomous actions
- • Internal platforms, CLIs, or harnesses other engineers use
- • Work in a setting where “the model said so” was not an acceptable answer
Benefits
- Primarily remote work arrangement
- Bonus
- Equity
Company Overview