Agentic AI Software Engineer

A four-person team inside an S&P 500 reinsurer is building autonomous agents that execute real underwriting and pricing decisions. Not internal tooling, not spend-management dashboards, not a chatbot for employees. Real financial judgment, at production accuracy, on real submissions.

It's already working. In under two years the team shipped chat, document extraction, deal pricing, and market comparison to almost 1,000 employees.. a few hundred thousand documents ingested and growing, other engineering teams consuming their pipeline. The COO ran the numbers, saw this team moving revenue, and pulled it out from under the CTO to report directly to him. His 2026 strategy is now this team's roadmap. The mandate isn't "keep going." It's "go faster."

Why the team stays small
No PM. No DevOps. No ML-only specialists. Everyone owns front-end, back-end, database, infrastructure. Not understaffing, a bet that a small team with real leverage outships a big one with layers. Real influence on revenue, not another engineer implementing someone else's architecture. That's the trade.

What's next, and undecided
Agentic orchestration. Role-specific sub-agents for underwriters, actuaries, and claims. Not built yet, architecture isn't settled. Today it's one agent, one big toolbox, no supervisor layer. The team runs experiment-led: one-week time-boxed builds to prove something sticks, then a real production rollout. If you have a real opinion on when multi-agent complexity earns its keep, that's the conversation, not a spec.

What you'd build
Agentic workflows that execute real underwriting decisions, not just answer questions. Role-specific sub-agents, from scratch. A client advocacy tool this year: internal CRM built from underwriters' notes, without buying Salesforce.

Stack
.NET/C# is the backbone, but the team isn't dogmatic. Python and TypeScript both run in production, whatever solves the problem best, with type safety as the one non-negotiable. React is a real gap here; strong React matters. AWS-native: Bedrock, Textract, Postgres/pgvector. OpenAI APIs, embeddings, RAG, custom agent orchestration.

The hard part isn't shipping agents. It's not breaking them. Preventing regression when a prompt or model changes, at a scale where "it worked in testing" isn't good enough.

Who fits

• Ships end to end without waiting for a ticket

• Strong software engineer, not an ML/data scientist by trade

• Comfortable in a room with underwriters and executives, translates what they actually need

• Entrepreneurial instinct: thinks about what this builds toward, not just what's in the spec

• 5–10 years shipping production systems

• Has genuinely put LLMs into a real workflow: tool use, multi-step orchestration, not a wrapper around an API

Growth here means creating leverage, not managing a team. Constant contact with the COO, not a manager relaying his priorities secondhand. Your work shows up in revenue, not in an internal tool nobody outside engineering ever sees.

Not for you if

• You want tickets and specs, not ambiguity

• You need a DevOps team backing you up

• You'd rather model data than ship product

100% remote. Quarterly travel. EST or GMT preferred.