Lago is the leading open-source billing platform for usage-based pricing. We power billing for companies like PayPal, CoreWeave, Mistral AI, and Synthesia — handling real-time metering, hybrid pricing models, credit wallets, and multi-entity invoicing. We raised $22M (Series A led by FirstMark, seed backed by Y Combinator) and are growing fast across AI/ML, developer infrastructure, and enterprise segments.
We’ve been building AI-powered internal agents that automate key GTM workflows — lead qualification, prospect research, outreach drafting, customer monitoring, and more. These agents run on Claude Code, connect to our CRM, Slack, and email, and have been iteratively refined with real production data over several months. They work. But right now they’re monolithic skills, operated by one person. We want to turn them into a production-grade multi-agent platform that the whole team can use — and that improves itself over time. You’ll work directly with Lago’s founders to own this transformation: from working prototypes to a scalable internal AI platform that becomes the backbone of our GTM operations.
Decompose monolithic AI skills into composable agents. Our current automations are single mega-prompts that handle enrichment, scoring, cross-referencing, drafting, and monitoring in one pass. You’ll break them into discrete, testable agents with clear inputs, outputs, and failure modes — built on Claude Code and the Claude Agent SDK.
Make the system multi-user. Today, one person triggers everything. You’ll design it so multiple team members can interact with the agents from their own context — different permissions, different views, different triggers — with a shared state layer that keeps everyone in sync.
Build self-improvement loops. These agents already incorporate learned corrections from real-world feedback. But today those corrections are manually written into prompts. You’ll build systems where feedback signals are automatically captured, calibration rules are proposed and reviewed, and the agents track their own accuracy over time.
Extend to new workflows. Lead qualification is the first use case, not the last. We have plans for deal-stage automation, customer intelligence, renewal risk scoring, competitive monitoring, and more. You’ll establish the patterns, tooling, and architecture that make spinning up new agents fast and reliable.
Ship, measure, iterate. This isn’t research. Every agent you build runs in production, handles edge cases gracefully, and earns the trust of the people using it daily. You’ll instrument everything and use the data to improve.
Strong software engineering fundamentals. You’ve built and shipped production systems. You’re comfortable with APIs, databases, async workflows, and CI/CD. Python or TypeScript preferred, but we care more about engineering taste than specific languages.
Hands-on with LLMs. You’ve gone beyond tutorials — maybe you’ve shipped a side project, contributed to an agent framework, or prototyped something real at work. You don’t need to have run eval pipelines at scale, but you should understand why you’d want one. You’re comfortable with tool-use APIs, structured output, and the unglamorous work of making AI reliable. Experience with Claude or Claude Code is a plus.
Agent architecture thinking. You understand the difference between a chain, a router, and a fully autonomous agent. You’ve thought about (or built) systems where multiple AI agents coordinate, share state, and handle handoffs. Familiarity with tool-use patterns, memory management, and human-in-the-loop design.
Product instinct. You won’t just build what’s specced — you’ll shadow the team, watch how workflows actually run, and identify where AI leverage is highest. You think in workflows, not features.
Comfort with ambiguity. This is a new function. There’s no team, no existing architecture, no playbook. You’ll define the stack, set the standards, and make the tradeoffs. The founders will give you context and access; you’ll figure out how to build it.
Nice to have
At 3 months: The first monolithic skill is decomposed into independent agents. At least two team members are using the system autonomously. Accuracy is instrumented and baselined. At 6 months: The self-improvement loop is live — feedback signals feed back into agent calibration with human-in-the-loop approval. At least one new workflow (beyond the first) is in production. At 12 months: The AI agent platform is a core part of Lago’s operations. Multiple team members rely on it daily. New agents can be spun up in days, not weeks. The system demonstrably outperforms its 3-month baseline.
Why this role is interesting You’ll work directly with Lago’s founders — the actual users of these systems — with extremely short feedback loops. You’re not starting from zero: you’re starting from working prototypes with months of real calibration data, and your job is to engineer them properly. You’ll be building AI infrastructure for a company whose core product is infrastructure, so the bar for quality is high. And you’ll shape a new function from scratch — the patterns you set will define how Lago builds internal AI for years.
Please Include:
We’ll send a take-home assignment after the screening call.