Agentic AI Engineer - II
Hyderabad, IndiaHybridFull-timeMid · 2-5 years
About this role
About the role
As an Agentic AI Engineer, you will sit at the intersection of applied AI research and production engineering. You will architect, build, and operate agent systems that use LLMs as core reasoning engines, integrate external tools and APIs, and operate reliably in dynamic, open-ended environments. You will translate business goals into agent specifications, design robust tool-use and memory subsystems, create evaluation and monitoring frameworks, and harden agents with safety guardrails and human-in-the-loop controls. This role requires strong software engineering discipline, deep practical experience with LLMs and agent frameworks, and a pragmatic understanding of cost, latency, and failure modes in production.
Key responsibilities
- Architect multi-agent systems
- Design orchestration layers and agent lifecycle management for single-agent and multi-agent workflows using frameworks such as LangGraph, AutoGen, CrewAI, or custom orchestration.
- Define communication protocols (event buses, message schemas, RPC), role definitions for agents (planner, executor, verifier, critic), and fault-tolerant patterns for agent restart, retry, and state reconciliation.
- Example outcome: a resilient orchestration that continues progress after partial failures and provides per-task provenance.
- Design and implement tool-use pipelines
- Implement tool integration patterns: function calling (structured API signatures), multi-call planning (MCP), browser automation (Playwright/Puppeteer agents), code interpreters (sandboxed execution), and external system adapters (CRM, ticketing, databases).
- Ensure secure credential handling, rate-limiting, transactional safety, and idempotency for side-effecting tools.
- Example outcome: a function-calling pipeline that maps LLM intent to safe API calls with rollback semantics and audit logs.
- Build memory and context systems
- Create hybrid memory architectures: short-term context windows (token/window management), episodic buffers (session-level state), and semantic long-term memory (vector stores, knowledge graphs) with TTL and relevance scoring.
- Implement context pruning, retrieval-augmented prompting (RAG), and memory consolidation/updating strategies to keep agents effective for long-horizon tasks.
- Suggested patterns: chunking + embedding, temporal decay, versioned memory snapshots, and human-curated memory curation workflows.
- Define and run evaluation frameworks
- Establish automated tests and benchmarks for agentic behavior: task completion, stepwise correctness, error propagation analysis, hallucination rates, cost-per-task (API tokens + infra), and latency percentiles.
- Build simulation testbeds or synthetic environments to run large-scale evaluation sweeps and A/B tests before production rollout.
- Example metric: maintain >95% safe-execution rate for critical workflows and reduce unnecessary external calls by 30% vs. baseline.
- Collaborate cross-functionally
- Partner with product managers to translate KPIs into agent acceptance criteria, with researchers to prototype novel reasoning methods, and with data engineers to instrument telemetry and build labeled datasets for feedback loops.
- Produce clear runbooks, API contracts, and handover docs for SRE and customer-facing teams.
- Establish guardrails and human-in-the-loop flows
- Design hierarchical safety: static policy filters, runtime monitors (shielding), confidence thresholds, step-level validators, and human escalation for ambiguous or high-risk actions.
- Implement auditing, explainability (provenance + chain-of-thought capture), and opt-in human review for learning updates (RLHF/online tuning).
- Compliance focus: PII handling, access control, and GDPR-style data governance where applicable.
- Monitor and iterate on deployed agents
- Instrument agents for observability: structured traces, event logs, prompt/response diffs, and custom metrics tracked in LangSmith, Weights & Biases, Datadog, or a bespoke stack.
- Run post-mortems for failures, implement automated rollback or canary deployments for agent updates, and continuously reduce false-positive/negative behaviors through targeted retraining or prompt adjustments.
- Research and adoption
- Continuously evaluate emerging tools, architectures, and safety methods in the agentic AI space and drive pragmatic adoption plans internally.
- Prototype new capabilities (e.g., tool chaining patterns, self-debugging agents, or task decomposition algorithms) and measure potential ROI.
Suggested tech and design choices
- Orchestration: LangGraph, AutoGen, CrewAI, custom state machines, message queues (Kafka/RabbitMQ)
- LLMs & tool orchestration: Hugging Face Transformers, OpenAI/Anthropic APIs, LangChain tool patterns
- Memory & retrieval: Pinecone/Chroma/Weaviate, FAISS, vector schemas with metadata and TTL
- Browser/automation tools: Playwright, Puppeteer, Selenium, Headless Chrome sandboxes
- Sandboxing & execution: Firejail, gVisor, containerized code runners, strict resource limits
- Observability: LangSmith, W&B, OpenTelemetry, Prometheus/Grafana, structured event stores
- Infra: Kubernetes, serverless functions, Redis, PostgreSQL, S3-compatible stores, Terraform
- Security: Vault for secrets, fine-grained IAM, request signing, audit logging
Operational expectations
- Own the lifecycle of agent features from prototype to production and support them in rotation for incident response.
- Ship measurable improvements every quarter: e.g., lower human escalation rate, improved cost-efficiency, fewer incidents.
- Mentor junior engineers, contribute to design reviews, and document patterns as internal best-practices.
Nice-to-haves
- Experience with causal/temporal reasoning methods, symbolic planners, or neuro-symbolic hybrids.
- Prior product experience shipping LLM-driven automation in customer-facing contexts.
- Publications or OSS contributions in agent frameworks, tool-use libraries, or safety tooling.