Senior Software Engineer, Agentic Platform - Moveworks

The Role

We're building the runtime infrastructure that powers Moveworks' AI agents — the systems that orchestrate, execute, and deliver agent responses to millions of enterprise users in real time. This role sits where large language models meet distributed systems. You'll apply machine learning, LLM, and distributed systems expertise to make agents plan, reason, and use tools reliably, and you'll build the production systems that run those agents at enterprise scale.

Our AI agents can plan, execute multi-step workflows, call tools, wait on human input, and resume — all while maintaining correctness, observability, and low latency. The systems that make this possible are what you'll build and own.

What you get to do in this role:

• Agent orchestration engine — A state machine that manages long-running agent sessions, coordinating planning, execution, and user interaction across multiple LLM calls and tool invocations
• Model routing and adaptation — Choose, route between, and fine-tune or distill LLMs for specific agent tasks. Meet enterprise requirements such as granular permissions, data freshness, explainability, and preserving general knowledge.
• Distributed session management — Lease-based ownership using conditional writes, heartbeat protocols, and crash recovery via checkpointing
• Event-driven message pipeline — FIFO queues for ordered delivery, consumers for event processing, and real-time streaming via gRPC and Socket.IO
• Structured concurrency — Running multiple concurrent tasks per session with fail-fast semantics and graceful cancellation
• Observability infrastructure — OpenTelemetry instrumentation, distributed trace context propagation across async boundaries, custom span lifecycle management for sessions that span minutes
• Caching and state layers — Redis, DynamoDB KV stores with per-org/per-bot scoping, batch read optimization, and hot-reload configuration
• Tool and agent connectivity for LLMs: Build MCP and agent-to-agent (A2A) capabilities, including tool discovery, namespacing, and tool selection over large tool catalogs, so the LLM picks and calls the right tool from thousands of candidates.
• Token efficiency and context engineering — Lower the agent's cost and latency through prompt and context discipline, caching, speculative and parallel model-call scheduling, and reducing tokens spent across multi-step LLM reasoning chains.
• Open-weights model adoption and evaluation — Quickly stand up newly released open-weights LLMs at different quantization levels, run them in a test environment, and benchmark their accuracy, latency and cost on agent tasks. Use the results to decide which models reach production.

Function: Engineering
Experience Level: Not Applicable