AI Engineer | Agent Engineer (Agentic AI)

Why PwC? 

What you’ll do: 

• • Build and extend our agent harness in TypeScript: the runtime loop, tool dispatch, streaming, retries and tool-call repair, state persistence and recovery, and the guardrails that keep agents reliable and on-budget. 

  • Design context and memory strategies for long-horizon work: context compaction and summarisation, working versus long-term memory, prompt and response caching, cache-aware prompt construction, and token budgeting across multi-step runs. 

  • Architect multi-agent and sub-agent systems: lead and worker patterns, task decomposition, parallel sub-agent execution, hand-off and delegation contracts, shared state, and result aggregation back to a lead agent. 

  • Design and expand agent tool ecosystems with well-typed, testable schemas: a sandboxed command tool giving agents a real filesystem and shell, file and document operations, code execution, web search, and connectors that surface enterprise and client data sources (for example MS Graph, SharePoint, Outlook and Teams) to agents as tools. 

  • Route across multiple models and providers through our LiteLLM-based gateway, tuning model selection, fallbacks and routing policy for quality, latency and cost - including the use of small language models for narrow, high-volume or latency-sensitive steps. 

  • Engineer retrieval and RAG pipelines that ground agent outputs in trusted sources: chunking, embeddings, vector search on PostgreSQL with pgvector, hybrid and re-ranked retrieval, and retrieval exposed to agents as tools. 

  • Produce structured, reliable model outputs using typed schemas (for example Zod with the Vercel AI SDK) and design robust prompts, system instructions and agent skills. 

  • Build evaluation harnesses and offline batch evals, including deep-research style runs and trajectory-level assessment of tool use, to measure accuracy, faithfulness, cost and latency - and use those signals to iterate. 

  • Instrument agents with observability and tracing (for example Langfuse, OpenTelemetry) and track token usage, cache hit rates and cost in production. 

  • Produce clean, maintainable, efficient code deployed at scale in the Azure cloud; scaffold new projects, pair with engineers, and review pull requests. 

  • Contribute to team stand-ups and the broader software development lifecycle, and participate in firmwide AI and engineering forums. 

What we’re looking for: Applicants must be able to demonstrate the following key capabilities. We do not expect every box ticked, strong fundamentals and a track record of shipping LLM-powered products matter most. 

Core skills: 

• • Strong TypeScript development experience, ideally in a Node.js backend context, with confident use of the type system to model tool schemas, agent state and structured outputs. 

  • Hands-on experience building agents with a modern agent framework or SDK, with the Vercel AI SDK strongly preferred (streaming, tool calling, structured generation, provider abstraction). 

  • Demonstrated experience building or operating an agent harness or runtime loop in production: tool dispatch, error handling and tool-call repair, interrupt and resume, and long-running task execution. 

  • Practical understanding of context engineering: context window management and compaction, memory design, prompt and response caching, and cost/latency trade-offs across a multi-step run. 

  • Experience designing multi-agent architectures, including sub-agent orchestration, delegation boundaries, shared state and failure handling. 

  • Deep, practical knowledge of prompt engineering, agentic patterns, tool and function calling, and structured outputs. 

  • Experience routing across multiple models and providers through a gateway such as LiteLLM, and reasoning clearly about model selection trade-offs. 

  • Experience building and optimising RAG and retrieval pipelines, including their evaluation, with industry-standard tooling. 

  • Strong experience with PostgreSQL, including pgvector for embeddings and hybrid or semantic search, and using Postgres as the durable store for agent state, memory and traces. 

  • An evaluation-first mindset: designing evals for agent behaviour, and using evidence rather than demos to reach defensible conclusions about quality. 

  • Strong critical thinking, analytical rigour and outstanding attention to detail. 

  • Proper source code management and confident use of Git, and comfort working in a large shared monorepo (for example Nx or Yarn workspaces) with project boundaries and shared libraries. 

  • Good written and verbal communication, and the ability to work effectively with remote teams. 

  • A proactive, problem-solving approach and the ability to solve complex problems as part of a team. 

Desirable skills: 

• • Experience with small language models (SLMs): selecting, fine-tuning, distilling, quantising or self-hosting smaller models, and using them for routing, classification, extraction or guardrail steps within a larger agent system. 

  • Experience with LLM observability and cost/quality tracing (for example Langfuse, OpenTelemetry). 

  • Experience designing sandboxed execution environments for agents: containerised or virtualised sandboxes, shell and filesystem access exposed as a command tool, isolation and resource limits, and safe tool surface design. Familiarity with emerging tool-interoperability standards such as MCP is a plus. 

  • Experience building connectors and integrations to enterprise systems and data sources, including authentication, permissioning, incremental sync and rate-limit handling. 

  • Full-stack TypeScript experience (NestJS, Next.js / React) with either tRPC or GraphQL / Apollo including code generation, and building interfaces for agent products such as streaming and human-in-the-loop review. 

  • Experience with computer-use or browser-automation agents (for example Playwright). 

  • Experience with document generation pipelines (DOCX, PPTX, PDF) or rich-text editor frameworks such as Lexical, ideally over a real-time collaborative layer (Yjs / Hocuspocus or equivalent CRDT tooling). 

  • Working Python capability, including building and running production services (for example FastAPI) as well as evaluation, data work and model experimentation; comfort moving across the TypeScript and Python stack is valued. 

  • A research background in ML or LLM development, and the ability to identify emerging techniques and apply them to practical situations. 

  • Knowledge of classical ML and deep learning, preferably in NLP. 

  • Experience with well-structured backend services at scale: microservices and containerisation (Docker), and ideally domain-driven and event-driven patterns such as domain / application / infrastructure layering, repositories and domain events. 

  • Experience with durable job queues and background processing for long-running agent workloads (for example BullMQ on Redis, Azure Service Bus), and with Redis for caching, locks and shared state. 

  • Experience developing on cloud environments, particularly Azure (Azure OpenAI, AI Search, Blob, Key Vault, App Insights). 

  • Knowledge of agile software development lifecycles (SDLC), and exposure to CI/CD pipelines (Azure DevOps, GitHub Actions) and infrastructure as code (Terraform).