Backend Engineer, Agent Infrastructure

Specter; New York (NY); On-site; USD 150,000 - 210,000; Full time

Job summary

Specter’s agents run unattended on customer data overnight, on schedule, and on demand. They read general ledgers, query banks, post journal entries, and send messages on behalf of finance teams. The infrastructure under them — orchestration, retries, observability, sandboxing, cost control — is the difference between a demo and a system a CFO trusts with the books.

This role builds that infrastructure. It's one of the highest-leverage engineering roles at the company, because everything else in

Qualifications

• Strong backend engineer with Python/PostgreSQL experience.
• Focus on correctness and recovery: idempotency and exactly-once vs at-least-once.
• Experience with LLM-shaped workloads and long-running tasks.

Responsibilities

• Build and own the agent runtime including scheduling, queuing, retries, idempotency and audit trails.
• Design an observable, replayable, auditable stack for agent actions.
• Optimize runtime efficiency: minimize token spend, latency, and manage parallelism.
• Create integration layer to connect agents to ERPs, banks, and spreadsheets across systems.
• Define production readiness criteria for agent code across the org.

Skills

Python

PostgreSQL

Async

Distributed systems

Debug production issues

Tools

Vector stores

Graph databases

Job description

The role

Specter's agents run unattended on customer data overnight, on schedule, and on demand. They read general ledgers, query banks, post journal entries, and send messages on behalf of finance teams. The infrastructure under them - orchestration, retries, observability, sandboxing, cost control - is the difference between a demo and a system a CFO trusts with the books.

This role builds that infrastructure. It's one of the highest-leverage engineering roles at the company, because everything else in the product runs on top of what you build.

What you'll do

• Build and own the agent runtime: scheduling, queueing, tool execution, retries, idempotency, audit trails, replay. Long-running, partially-failing, non-deterministic workloads are the default, not the edge case.
• Design the observability stack. Every action an agent takes should be inspectable, replayable, and explainable to a non-technical reviewer six months later during an audit.
• Make agent runs cheap and fast. Token spend, latency, and parallelism all matter, and they trade off against each other in non-obvious ways.
• Build the integration layer that connects agents to ERPs, banks, spreadsheets, and the dozens of source systems customers actually use. Each one is different. Most of them are old. Many are inconsistent, under-documented, rate-limited, or shaped by customer-specific configuration.
• Set the standard for production reliability across the engineering org. You're the person who decides what "ready to ship" means for agent code.

What we look for

• Strong backend engineer. Python, Postgres, async, queues, distributed systems. You've debugged something that only fails in production at 3am, and you can talk about what you learned.
• You think in terms of correctness and recovery, not just happy paths. Idempotency, exactly-once, at-least-once - you know the difference and when each matters.
• Comfort with LLM-shaped workloads: non-deterministic outputs, partial failures, long-running tasks, retries that have to be careful about side effects.

Nice to have

• Experience with vector stores and graph databases.
• Prior work on agentic systems and ETL.
• Experience with banking, ERP, or other integrations related to the finance stack.