Senior AI Engineer – Harness Engineering (Kimchi)

Kimchi is the AI platform inside CAST AI that helps companies run LLMs on their own Kubernetes clusters and builds the execution layer where agents do real work. The current infrastructure supports multi-model inference with intelligent routing and deployment flexibility across GPUs and VPCs. The hired engineer will work on coding agents, agent runtimes, orchestration systems, and reliability engineering to ensure task completion.

• Experience using AI coding agents in real work scenarios and familiarity with agents such as Claude Code, Codex, OpenCode, and Cursor.
• Strong production experience with TypeScript and Go, comfortable moving between both languages.
• Ability to think in harness engineering terms and diagnose agent failures effectively.
• Capable of driving features end-to-end from design through iteration without reliance on layers absorbing ambiguity.
• Knowledge of Kubernetes, cloud providers (AWS, GCP, Azure), and observability tools like Prometheus, Grafana, and Loki.
• Experience with CI/CD tools such as GitLab CI and ArgoCD.

• Build and evolve the agent harness by shipping hooks, permission gates, and context compaction based on diagnosed failures.
• Own long-horizon execution with spec-driven prompting, worktree-per-slice git strategy, Ralph loop recovery, and stuck detection.
• Architect planner, executor, and evaluator pipelines using different models for planning, execution, and evaluation.
• Manage agent memory and context persistence across sessions to prevent context rot.
• Own the harness surface including CLI, TUI, MCP integrations, sandboxed execution, and telemetry using TypeScript on the surface and Go at the core.
• Engineer solutions to prevent repeated agent mistakes by enforcing constraints like pre-commit lint checks and permission gates.
• Separate generation from evaluation to improve task grading and reliability.
• Transition agents off laptops to infrastructure with persistent state, context management, compaction, and tool-call offloading.