Full Stack AI Engineer Associate Manager

About Accenture Data & AI The beginning of a new Data & AI decade that will reshape work and society is underway. Accenture is stepping boldly into this future with a clear strategy and purpose: to help clients optimise and reinvent their businesses with data and AI — backed by a $3 billion investment and a commitment to industry-defining work. With over 45,000 professionals dedicated to Data & AI, Accenture's Data & AI organisation brings together Experienced Innovation, Strategic Investment, Exceptional Talent, and a Power Ecosystem to deliver outcomes at the frontier of what is possible. About the Role Accenture is establishing a dedicated Agentic AI Ninja Team — a group of highly experienced engineers tasked with solving the most complex challenges at the frontier of autonomous AI. This is a technical leadership position for engineers who have designed, built, and operated production-grade agentic systems at enterprise scale. The Associate Manager will be responsible for the full delivery lifecycle of agentic AI applications: from system architecture and agent design through to deployment, evaluation, and production observability. Candidates will be expected to bring deep hands-on experience with autonomous agent frameworks, multi-agent orchestration, advanced retrieval architectures, and enterprise-grade integration — not proof-of-concept or prototype experience, but systems that have operated under real business conditions with real consequences. This role also carries a technical leadership responsibility: guiding engineers, setting delivery standards, and owning the quality of output across fast-moving, high-visibility client engagements within Accenture's Data & AI practice. Position Responsibilities Agentic System Design and Delivery Architect and deliver production-grade autonomous AI systems — agents that plan, reason, invoke tools, recover from failures, and integrate with enterprise backends across cloud platforms. Select and apply appropriate reasoning patterns (ReAct, Chain-of-Thought, Tree-of-Thought, Plan-and-Execute, Reflexion) based on task complexity, latency requirements, and verifiability needs. Author structured agent specifications using spec-driven development practices; apply AI-assisted engineering tooling (Claude Code, Codex) to accelerate delivery without compromising rigour. Design and maintain prompt architecture for production agents — system prompt structure, few-shot example design, structured output schemas, prompt versioning, and A/B testing of prompt changes as production artefacts. Agent Harness and Orchestration Design and implement the agent harness: agent instantiation, persona and instruction loading, tool binding, memory initialisation, and lifecycle management from invocation to termination. Architect multi-agent orchestration topologies — supervisor/worker hierarchies, event-driven graphs, parallel execution — with defined A2A handoff contracts, shared state schemas, and structured escalation paths. Configure the LLM gateway and model routing layer — directing agent calls by task type, latency, cost, and capability — using provider-agnostic abstraction (LiteLLM or equivalent) across LLM providers. Tool Layer, Context, and Memory Design, build, and maintain MCP servers exposing enterprise systems, APIs, databases, and SaaS platforms as agent-accessible tools — with robust schema design, error handling, idempotency, and retry logic. Translate business processes into agent-executable skills, structured instructions, and reusable workflows — bridging the gap between business requirements and agent implementation. Build context engineering pipelines — assembling the right information into the agent context window across multi-turn and long-running tasks, with explicit management of context budget and retrieval triggers. Implement memory architectures — episodic, working, and long-term — using appropriate backends (vector stores, relational databases, cache layers) matched to each agent use case. Knowledge Layer and Engineering Design RAG pipelines for agentic contexts: hybrid search, semantic re-ranking, late chunking, multi-vector retrieval, and metadata filtering; manage the full lifecycle from ingestion through quality evaluation. Build MCP-connected knowledge sources exposing structured and unstructured data assets as governed, agent-accessible tools. Implement Text-to-SQL capabilities — prompt-to-query translation, schema grounding, query validation, and safe execution against live enterprise databases. Integrate Elasticsearch as a retrieval backend: full-text search, BM25 scoring, faceted filtering, and hybrid semantic-lexical strategies. Design knowledge graph and ontology layers providing agents with structured representations of domain entities and relationships for precise reasoning over interconnected enterprise knowledge. Agent Ops, Registry, and Observability Operate and maintain production agentic systems using AgentOps, LLMOps, and DevOps practices — CI/CD pipelines for agent code and prompt changes, automated evaluation gates, and deployment strategies (blue/green, canary) across environments. Manage an agent and asset registry — versioned catalogue of agents, tools, skills, prompts, and workflows — enabling reuse, governance, and controlled promotion across development, staging, and production. Define and implement agent evaluation frameworks: golden dataset construction, LLM-as-judge pipelines, tool-call accuracy measurement, trajectory evaluation, and faithfulness scoring. Build agent testing suites distinct from evals — unit testing agents with mocked tools, integration testing multi-agent handoffs, and simulation environments for pre-production scenario testing. Design HITL feedback capture: structuring human corrections and approvals as refinement signal for continuous improvement. Build production observability from day one — distributed tracing, token-level cost tr