About Traversal

Traversal is the AI Site Reliability Engineer (AI SRE) for the enterprise.

Production complexity was already outpacing what engineering teams could manage manually, and AI-generated code is accelerating that gap. Traversal is built for that challenge, autonomously understanding and reasoning across even the largest, most complex production environments to diagnose, fix, and prevent incidents. Our mission is to free engineers from endless firefighting and give them more time to focus on creative, high-impact work.

Today, Traversal operates in mission-critical environments at some of the world’s largest enterprises, including PepsiCo, Nubank, Harvey, and DigitalOcean. Our roots remain deeply embedded in AI research, and we’ve brought together researchers from institutions including MIT, Harvard, Berkeley, Columbia, and Cornell with world-class technical staff and operators from companies like Google, Meta, Datadog, ServiceNow, and Citadel Securities to take on one of the hardest problems for AI to solve. Traversal is backed by Sequoia Capital, Kleiner Perkins, Redpoint Ventures, and Khosla Ventures.

The Role

Our core product is our AI SRE agent. As a Product Engineer on the Agents team, you'll own how that agent behaves, where it shows up, and whether it holds up under real production load. That means working end-to-end on the harness and context engineering that shape agent reasoning, the streaming and orchestration layers that put it in front of engineers mid-incident, and the observability and evaluation systems that tell us whether any of it is actually working.

Three areas drive impact on this team:

Agent Behavior: Improving how the agent reasons, decides, and communicates through harness design and context engineering. This is prompt and tool architecture, not prompt tweaking: what the agent sees, when it acts, what it does with ambiguity, and how we measure whether a change made it better.

Live Surface Area: Getting agent intelligence into the places engineers already work, in real time. Slack channels, the web app, and whatever comes next. Streaming, orchestration, and long-running work that survives restarts and reconnects.

Stability at Scale: Our agents run for hours against noisy, high-volume customer telemetry. Making that reliable means observability into agent behavior, scalable data patterns, and systems that degrade gracefully instead of silently.

Each of these requires product vision and autonomy. You'll be deciding what to build as often as you're building it.

Responsibilities

Agent Development: Build AI agents with a mind for user experience and capabilities like long-horizon work, proactiveness, resilience, readability, and evidence citation.

User Experience: Innovate and drive the evolution of AI UX/UI design, rapidly iterating based on user feedback to improve overall usability.

Harness and Context Engineering: Design the tool interfaces, context assembly, and decision logic that determine how well the agent performs on real incidents.

Evaluation: Build the eval harnesses and offline datasets that let us ship agent changes with confidence rather than vibes.

Live Delivery: Design and implement the streaming, orchestration, and API layers that carry agent output into Slack and the product in real time.

Data and Infrastructure: Own efficient storage, retrieval, and processing across PostgreSQL, Redis, Kafka, and S3, in support of agents operating on large-scale time-series and topological data.

Product Impact: Translate agent capability into features that reduce cognitive load for on-call engineers, iterating quickly on user feedback.

Requirements

3+ years of software engineering experience, with a strong focus on full-stack or backend systems.

Hands-on experience building with agentic AI systems or LLM-powered products, including familiarity with tool use, context management, and the failure modes that come with both.

Strong experience with Python and web frameworks such as FastAPI.

Comfort working in a React and TypeScript codebase, enough to ship a feature end to end without waiting on someone else for the frontend.