Build agents that defend companies

We're building AI agents and the infrastructure around them to help security teams detect threats, investigate alerts, and respond. Our agents work across live security tools and data, write and tune Python detections, and bring humans in when judgment or approval matters. Teams at Ramp, EliseAI, and ClickHouse use Cotool.

We're looking for a generalist software engineer who wants to build the systems that turn improving models into more capable agents and ship the product around them.

What you'll build

• Agent harnesses. Improve the execution loop: tool selection, state management, retries, failure recovery, and the controls that let agents take useful actions safely.
• Context layers and memory. Get the right evidence, environment context, and prior knowledge into the right turn. Work on retrieval, context assembly, and compaction so long investigations stay coherent without blowing up latency or token cost.
• Evals and feedback loops. Turn real failures into reproducible cases, inspect traces, and measure whether a change actually makes the agent better.
• Tools and data infrastructure. Give agents useful interfaces to security systems and fast access to data through ClickHouse, APIs, and integrations.
• The product around the agent. Build interfaces for investigating findings, editing detections, inspecting evidence, and approving actions. Own features across the UI, backend, and runtime.

A project might start with an agent losing a critical fact halfway through an investigation. You'd follow the trace, change how context is assembled or compacted, build an eval that catches the failure, and ship the fix. Another might start with a customer workflow and end with a new tool, execution path, and interface.

Who will enjoy this

You build things because you want them to exist. You've pushed coding agents or your own agent projects far enough to develop opinions about where they break. You're excited to test those opinions against production behavior.

You have strong software fundamentals, move comfortably across unfamiliar parts of a codebase, and care about the details between a compelling demo and a system someone can depend on. You use AI tools aggressively and can explain, debug, and take responsibility for the code you ship.

Show us something you've built: a product, an open-source contribution, an agent experiment, or a technically ambitious side project. We're interested in your decisions, what failed, and what you learned.

The stack and team

TypeScript, React, Node.js/Express, Python, PostgreSQL, ClickHouse, and Redis. We use AI SDK for model integration and Braintrust for tracing and observability.

We're a small team in San Francisco backed by Andreessen Horowitz, WndrCo, Y Combinator, and Homebrew. You'll work closely with the founders, own substantial pieces of the product, and see your work used by security teams. We offer competitive salary and equity, healthcare, dental and vision coverage, and an unlimited token budget.

If you keep thinking about what agents could do with a better harness, better context, and better tools, we'd like to build with you.