We shipped our first agent, Chloe, in 2026. The Agents team owns the platform underneath: the agent core, the eval and observability layer, the MCP surface that external agents (Claude, ChatGPT, ElevenLabs, n8n) operate Close through, and the orchestration that ties it together.
The team is running three major streams right now:
Voice Agents, the
LLM-powered chat assistant, and
Custom Agents. Most of the engineering problems here didn't exist a year ago. We're figuring them out in code, in production, in front of 11,000 paying customers.
One thing to know up front: we do move people between teams as the work shifts. Most engineers here end up on more than one team over their time at Close - this team is where you'll start, but over time you'll likely have the opportunity to work on many different projects.
This role is open at multiple levels - Software Engineer, Senior, and Staff. You don't need to pick one when you apply: we'll calibrate together during the process.
You are- A seasoned engineer. Python is our backbone, but perhaps you've worked across Go, Rust, or TypeScript. You pick the right tool for the workload rather than retreating to what you know.
- AI-native in production. You've shipped meaningful, impactful agentic features to users. You have opinions on retrieval, evals, tool design, context engineering, and where the current frontier models fall over.
- Working with AI in your day-to-day. You use AI tools in your own workflow to ship faster, write tighter code, and reason about unfamiliar parts of the codebase. You have a POV on where they help and where they get in the way.
- A builder first. You'd rather get a sloppy v1 in front of fifty customers than spend three weeks on abstractions. You ship.
- Battle-tested. You've debugged incidents where latency budgets didn't hold, owned a system everyone else relied on, or carried a pager for something with real customer impact.
- Close to the research. You read papers, or you've spent serious time in retrieval, RAG, RLHF, or fine-tuning. Not a researcher, but you can read one and tell us whether the result matters for our problem.
- Comfortable with non-determinism. Much of your output is probabilistic. Conventional patterns don't hold. You find this fun.
You will- Ship code execution for the assistant. The assistant decides when writing code beats answering non-deterministically - today that's generating charts and tables in Python on the fly; next it's reusable user-defined tools and calling external APIs to pull in whatever data the task needs.
- Build the eval and observability layer that tells us when an agent is getting better. Unit tests don't cut it for non-deterministic output. We run evals and tracing (LangFuse and our own tooling) as the bar for shipping: if we can't measure it, we don't ship it.
- Push generated UI forward. The backend increasingly decides what the user sees - the LLM picks the right presentation (table, chart, widget) and renders it in the assistant, with a full-screen experience and stored, referenceable artifacts on the roadmap. Effectively: customers generate the reports we used to hand-build, one custom report at a time.
- Take Custom Agents from prototype to GA. Event-driven agents that act on what's happening inside the CRM in real time - an email lands, an agent drafts the reply from knowledge sources and context, the user approves. This is where we differentiate from the general-purpose assistants: we see the events, we have the context.
- Make deterministic and non-deterministic systems work together. Sales processes need steps that happen every single time; LLMs are bad at that. You'll help fuse our Workflows engine with agentic steps so customers get reliability where it matters and intelligence where it helps.
- Handle the edges that make agents trustworthy. What happens to a fleet of running agents when a customer's AI credits run out? Pause semantics, recovery, and making sure nothing places a hundred calls that were supposed to happen last week.
- Pick the right model for the job. We use many providers, test new models constantly, and are moving toward cost-aware routing - simple summarization jobs shouldn't run on frontier-priced models. You'll call when something is production-ready and when it's still a demo.
Tech you'll touch: Python, Temporal, LangFuse, Pydantic, ElevenLabs, MCP, PostgreSQL, MongoDB - plus whichever LLM ships next.
