About Cerby At Cerby, security is everyone's business. Cerby extends your security perimeter to include your disconnected applications, providing identity automation and eliminating manual processes. It's one of the biggest blind spots in enterprise security, and we're the team closing it. If you want to work alongside the team solving one of the hardest, most overlooked problems in security, you're in the right place. Founded in 2020, Cerby is backed by leading investors and trusted by enterprises across the globe. OVERVIEW & IMPACT Every application Cerby protects needs Cerby to know how it works — its API, or the clicks behind its login, password rotation, MFA setup, or access change — and that knowledge breaks whenever a vendor redesigns its site or changes its API. We are looking for a Staff AI Engineer to own the AI agents that onboard new applications and repair automations when they drift. You will start with two agent systems we have already built and benchmarked: API catalog generation: reads a vendor’s public API docs in a headless browser and produces a validated OpenAPI spec, with per-job cost ceilings and an accuracy benchmark. Browser workflow generation: performs a task in a SaaS app’s web UI and writes a replayable automation workflow, handling MFA, one-time codes, and CAPTCHAs along the way. Your job is to take them to production, make them self-healing, and set the direction for agentic automation at Cerby. These agents handle real customer credentials and browse untrusted sites, so security is part of the job, not an afterthought. KEY RESPONSIBILITIES Ownership and Production • Own agentic onboarding and self-healing end to end: architecture, roadmap, and on-call once live. • Take the existing systems to production on durable workflows (Temporal), with the observability to run them. Self-Healing Automation • Detect drift — a changed login page, a moved button, a renamed API field — from failed runs, before customers notice. • Build agents that diagnose the break, repair the workflow or spec, and re-validate the fix against real runs before it goes back into service. • Decide what heals automatically and what needs human review. Evaluation and Cost • Own the benchmarks for generated specs, workflows, and repairs, and make “good enough to ship” a measured threshold. • Own cost and latency per output, choosing between frontier and open-weight models on measured accuracy versus cost. Security • Own the threat model: prompt injection from page content, credential isolation, token handling, sandboxed browsing, and egress. • Keep credentials out of prompts, logs, queues, and stored run state. Collaboration • Partner with the integrations, automation-runtime, and platform teams; write the design docs they build on, and mentor contributors. QUALIFICATIONS & REQUIRED SKILLS Experience • You have shipped an LLM agent with tool use that ran unattended for real users, and owned a backend system in production. Technical Skills • Strong in Python or TypeScript, and productive in both. • LLM agents and tool use: agent loops, structured outputs, retries, and the failure modes they create. • Eval discipline: you have built benchmarks or eval harnesses and used them to make decisions. • Depth in browser automation (Playwright or similar) or backend and API-integration systems, and comfort in the other. • Distributed systems: queues, idempotency, and durable workflow engines (Temporal, Step Functions, or similar). • Security judgment for systems that handle secrets, including agent-specific threats such as prompt injection. Desirable • Temporal in production; Pydantic AI or a comparable agent framework. • Open-weight model serving (Fireworks, Together, Bedrock) and LLM cost modeling. • Computer-use agents or browser extensions (Chrome MV3); OpenAPI tooling. • Identity, IAM, or credential-management experience; AWS and Terraform. Soft Skills • Clear technical writing and trade-offs other teams can depend on. • You take an early system to production without waiting to be told how, and care that it keeps working after launch. WHAT SUCCESS LOOKS LIKE • 6 months: both systems run in production for real workloads, with improving error rates and cost per output, and drift is detected automatically. • 12 months: onboarding a new application is predominantly automated, and most drift-driven breakages heal without a human.