AI Harness Engineer

Audria, Noida (on-site, five days a week), full-time.About AudriaAudria is an agentic AI notetaker for teams that work from an office. It converts in-person conversations, including standups, whiteboard sessions, and desk-side reviews, into completed actions in Gmail, Slack, Jira, Linear, and Confluence. It identifies team members by voice and adds relevant context to tools such as Claude and ChatGPT.Role OverviewThis role exists to build Audria's agent harness, the system that wraps around the model and determines whether it operates reliably. The work spans four connected layers: prompt design, context engineering (what the model sees at each step), harness engineering (the tools, guardrails, and environment the agent operates in), and loop engineering (how the agent observes, acts, verifies, and recovers over many steps). Harness and loop engineering are the center of this role, not prompting alone. This work directly shapes how Audria turns conversations into completed actions across tools like Gmail, Slack, Jira, Linear, and Confluence.Responsibilities- Design and build Audria's agent loop: the repeating cycle of model calls, tool execution, and decisions about when to continue or stop, sometimes called loop engineering- Engineer context: decide what goes into the context window at each step, including retrieved data, tool results, message history, and MCP resources, and manage compaction as sessions grow long- Engineer for caching: structure prompts and context so that prefix and KV cache hit rates stay high, keeping latency and inference cost down as agent sessions get longer- Design the tools, guardrails, and validation logic that keep the agent correct and safe when running with minimal supervision- Build backend systems using FastAPI, with scalable databases, caching layers, and containerized deployments, to run the harness in production- Set up observability and evaluation infrastructure, using tools such as Langfuse or comparable platforms, to trace every model call and tool call and measure whether the harness is actually working- Approach problems end to end: evaluate approaches, benchmark them against alternatives, and be upfront about caveats and limitations- Communicate findings, tradeoffs, and technical decisions clearly, in writing and in conversation- Stay current with AI developments, including new models, harness patterns, and agent tooling, and bring relevant advances into Audria's workRequirements- Experience building or working deeply with agent harnesses: the execution environment around an LLM, including tools, context management, the agent loop, and guardrails- Practical understanding of context engineering, including what should go into a model's context window at each step and how to manage it as a session grows- Practical understanding of prompt or KV caching and how context design affects latency and inference cost- Experience building systems with FastAPI, scalable databases, caching, and Docker containers- Ability to evaluate and benchmark AI systems, including understanding their limitations- Strong first-principles problem solving skills- Excellent communication skills, written and verbal- Genuine enthusiasm for AI, with active awareness of current developments, state of the art models, and agent toolingPreferred Qualifications- Active engagement with open agent harness projects such as OpenClaw, Hermes Agent, or OpenCode, including reading or contributing to their code- Experience with LLM observability and eval tooling such as Langfuse- Extensive hands-on use of Claude in highly optimised systems or workflowsAdditional Information- Location: Noida, India (on-site, five days a week)- Employment type: Full-time- Compensation: Based on experience- Reports to: Founding teamHow to ApplyCandidates should submit a resume, along with links to relevant projects, code, or research, where available.