AI Software Engineer, Agent Harness

The Opportunity

We serve open-weight models and our own bespoke checkpoints on EnCharge hardware. The models change often, and the harness around them needs to keep up. You own this layer that runs agents against files, tools, documents with permissions, memory, unattended execution, and real outputs. It will be assembled from a combination of open-source and bespoke code.

Key Responsibilities

• Own the harness architecture end to end — agent loop, safe execution, context management, knowledge base, memory, permissions, orchestration, outputs, interfaces, observability — one component per layer, with clear interfaces so layers can be swapped.
• Build the pieces with no open-source equivalent e.g. session semantics, enforced permissions, memory in a human-editable file, orchestrator, and outputs.
• Keep pace with the models: adapters, prompt formats, tool-call schemas, stop conditions, benchmarking and evaluation.
• Make tool use reliable across models of uneven tool-calling quality — validation, repair, retries, fallbacks.
• Develop agents, tools, and MCP servers for internal and customer use cases, and review them for security before they ship.
• Build the evaluation harness: task suites, regression runs on every model or harness change, cost and latency per task alongside quality.
• Define the interfaces: session API, CLI, GUI, and an endpoint existing tools can point at.

Qualifications

• 10+ years of software engineering experience in backend systems or ML infrastructure
• Strong Python and at least one systems language (e.g., Go, Rust, C++)
• Have shipped and operated an agent loop in production — tool use, multi-step workflows, unattended runs
• Hands-on with RAG, context management, and memory for LLM applications
• Experience with sandboxing, isolation, and permission models for automated systems
• Have run open-weight models yourself and understand how quantization and serving choices change model behavior
• Comfort in fast-moving, ambiguous environments where you define the roadmap; strong product instincts

Nice to Have

• Contributions to open-source agent harnesses or coding agents
• Experience with agent benchmarks (e.g., SWE-bench, Terminal-Bench) and building internal task suites
• MoE serving familiarity e.g. expert placement, tensor parallelism, quantization etc.
• Observability for LLM systems
• Document parsing and indexing pipelines
• Desktop or GUI application experience