Voice/Conversational AI Engineer at SUPERAGENT | Jobyssey

Superagent AI is building the AI-native distribution layer for insurance - a $7T industry still run on manual workflows and legacy software. We're an early-stage, fast-moving team reimagining how insurance is sold, serviced, and supported, with AI at the core. Backed by top-tier VCs, we're an international team of exited founders, builders, and technologists.

Already trusted by 100+ agencies, the job now is to build the next version of the platform quickly and run it reliably as we continue to scale rapidly — a real-time voice platform, an agent/skills layer, omnichannel messaging, and deep insurance system integrations.

Why this role exists
Our AI voice agents talk to real policyholders every day, and they're already working. What they say, and how reliably they say the right thing, is the product, and we've built real strength here. Now we want to move faster and take it to the next level.

We're hiring a Voice/Conversational AI Engineer to add firepower to how our agents are prompted, evaluated, and improved. You'll design and scale the prompts and agent behaviors behind live calls, mine real call data for the next set of wins, and build the evals and benchmarks that let us push quality forward with speed and confidence. The goal is leverage: ship more improvements, faster, each one backed by evidence rather than guesswork.

This is a hands-on builder role at the intersection of applied AI, data, and product quality — an opportunity to raise an already strong voice platform to a new level.

What you'll own
Prompt engineering at scale — design, structure, and version the prompts and agent behaviors that drive live voice calls, in a way that stays maintainable as the number of agents and use cases grows.
Reliability of agent behavior — make the same prompt behave consistently across calls, edge cases, and model updates; catch regressions before customers do.
Call analysis — dig into real transcripts and call data to find failure patterns, missed intents, and opportunities, and turn those findings into concrete improvements.
Evals & benchmarks — build the eval harnesses, datasets, and benchmarks that measure agent quality (accuracy, task completion, latency, tone), and make them part of how we ship.
Data-driven decisions — set the metrics, run the experiments, and bring evidence to every change, so we push quality forward with speed and confidence.
Velocity through AI-native engineering — you use Claude Code and Codex every day to orchestrate coding agents and raise your leverage, without letting speed erode the reliability bar.

About you (must-haves)
Strong applied-LLM experience — you've built and shipped LLM-powered features in production, not just prototypes.
Serious prompt-engineering craft — you can decompose a hard task into reliable prompts/agent steps, and you have opinions on structure, versioning, and keeping prompts maintainable at scale.
Evals-first mindset — you've built (or badly wished you'd built) eval harnesses, benchmarks, and test sets, and you believe quality is something you measure, not vibe-check.
Comfortable in data — you can pull, slice, and analyze call/transcript data (SQL and Python/TypeScript) to find signal and quantify impact.
Solid engineering foundation — you write production code (TypeScript/Node preferred) and can integrate your work into a real backend, not hand it off as a notebook.
Rigor over intuition — you default to experiments and evidence, and you make the case with data.
You live in Claude Code / Codex (or equivalent) — fluent and opinionated about AI-assisted engineering. In 2026, we treat this as table stakes.
Ownership mindset — you own agent quality end-to-end, including in production, and you stay close to how it affects real agencies and their customers.

Bonus
Real-time voice/telephony context (Twilio, SIP, ASR/TTS, latency-sensitive audio) and how it shapes prompting and evals.
Building agentic systems in production — tools/skills frameworks, MCP, guardrails.
Fine-tuning, distillation, or eval-driven development at a deeper level.
Experience building LLM-as-judge or automated grading pipelines.
Insurance, fintech, or another regulated/compliance-heavy vertical.

Our stack & tools
Backend: TypeScript, Node, NestJS · monorepo (libs + apps)
Voice: real-time streaming audio, ASR/TTS, telephony (Twilio / SIP)
Data: PostgreSQL on AWS RDS (multi-tenant, row-level security)
AI: Anthropic + OpenAI models, prompt/agent orchestration, evals; fine-tuned/distilled models on our own data
Frontend: React / Next.js
Platform: AWS · Clerk (auth) · Datadog (observability) · Pylon + HubSpot (CS/CRM)
How we ship: Linear (with agent delegation), GitHub, and Claude Code + Codex as part of daily engineering

How we build (engineering principles)
Reliability is a feature. We don't ship what we can't keep running. The quality bar is non-negotiable.
AI-native by default. Claude Code, Codex, and coding agents are part of how we work, not a novelty. We expect leverage from them, every day.
Close to the customer. We prioritize real agency calls and feedback, not opinions in a room.
Ship small, ship often. Short cycles, fast feedback, evals over guesswork.
Ownership. Engineers own their work end-to-end — including in production.

How we work
Fast-moving distributed team with an AI-native engineering culture. We're closer to our customers than most competitors — that's our edge, and it drives how we prioritize. We ship things that work for real agencies, not demos.

Explore more jobs