> Markdown version of [/jobs/ext/4636644-software-engineer-full-stack](https://www.wearedevelopers.com/jobs/ext/4636644-software-engineer-full-stack). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md). --- # Software Engineer, Full Stack - **Company:** ANDE Corporation - **Location:** New York, NY, United States - **Experience:** Experienced - **Contract:** Permanent contract - **Skills:** Application Programming Interfaces (APIs), Artificial Intelligence, Amazon Web Services, Distributed Systems, Software Product Management, Recommender Systems, OpenAI, TypeScript, Large Language Models, Concurrency, AI Coding Agents, Backend, Information Technology, AWS Fargate, Amazon Simple Queue Service (SQS) - **Published:** November 8, 2026 - **Apply:** https://www.juju.com/job/28_25303_eaaf3e08c85e4351b59a82251c8a134b ## About the Role * A degree in Computer Science or Engineering and 2+ years building production software, including shipping LLM agents to real customers (tool use, multi-step planning, structured outputs, long-running conversations). * Experience owning evals in production: writing them, finding failure modes in real traffic, and improving quality turn by turn. * Strong TypeScript and backend systems skills: queues, concurrency, idempotency, and the failure modes of distributed systems. * Judgment about where a model belongs and where deterministic code does, and the discipline to make agent actions safe, auditable and reversible. * Daily fluency with AI coding agents. Our codebase is written to be worked on by them, and we expect you to raise that bar. * The ability to explain agent behavior and trade-offs to product, operations and customers., * You've been a founding or early engineer and know how to balance craft, ownership and speed. * You've built agent harnesses, orchestration layers or tool brokers, not just prompts. * Experience with search, ranking or recommendation systems. * Experience with marketplaces, booking or payments, where a wrong action costs money. * Experience defending agents against prompt injection from untrusted content. AWS (Bedrock, Fargate, SQS) and the OpenAI and Anthropic APIs. This is an in-office position (4 days/week) in New York City ## Description You will own the intelligence behind the Ande Agent. That's the concierge and its team of subagents that take a request from one sentence ("dinner for 12 with the Acme team next Thursday, somewhere private") to a confirmed, paid, on-policy booking. Along the way they source venues, compare options, email the venue and negotiate terms, and handle every change after that. This is not a chatbot or a demo. The agent acts on real money, on behalf of real executives, with real venues on the other end of the thread. It runs in production today for reservations and private dining. Our strategy depends on it carrying more of the work with every request: more kinds of entertainment, more cities, more of the workflow, without growing headcount to match. You will set how we get there., * The agent system. A root concierge plus specialist subagents for sourcing, comparison and venue coordination, running on a self-hosted agent harness behind queued turn workers. You will own its architecture, model choices and upgrade path, and how it holds up under latency budgets, provider failover and load. * Tools the agent can trust. About 75 typed capabilities that let the agent search, hold, book, pay and message. Each runs as the acting user, respects who they act for (an assistant booking for an executive), and turns consequential actions into proposals the booker confirms. You decide what the agent gets to do, and how safely. * Evals that gate every change. Scenario suites, graders and LLM judges calibrated against human labels, replayed from recorded runs on every PR and run in full nightly. Make them trustworthy enough that a model upgrade or prompt change ships on the strength of the number, and cheap enough to run all the time. * Venue correspondence. The agent negotiates with venues over email: it reads replies, pulls out offers, minimums, deadlines and payment requests, and knows when to push, when to wait and when to hand off to a person. All of it has to be resistant to anything a third party writes into a thread. * Memory and the supply learning loop. What the agent learns about a company, its people and its venues, with clear rules on what is shared and what stays private. And the loop that turns every venue conversation into structured supply data (which room actually fits 14, what a venue will do on a Tuesday) that makes the next recommendation better. * Fast judgments. Typed classifiers for consent, triage, ranking and urgency, each with labelled thresholds, a timeout and a fallback, rolled out from shadow to live. * The bar. You set the standard for how we build with AI: agent behavior as versioned config, product people proposing changes through evals, and AI coding agents as the default way the team ships. You mentor engineers and partner closely with our AI product manager.