AI Agent job map: turn job descriptions into a skills plan

How to read an Agent job description

Convert every keyword into three questions: what business result will you deliver, which layer owns the work, and what evidence proves that you have done it?

• “Agent planning and tool calling” implies state, contracts, loops and failure branches.
• “Enterprise knowledge base and RAG” implies parsing, retrieval, reranking, citations and updates.
• “LLM application delivery” implies business APIs, permissions, monitoring and cost.
• “Agent platform” implies orchestration, models, tools, tenants, capacity and governance.

Role map

• Agent application: a business assistant with APIs, prompts, tools and RAG.
• RAG / knowledge engineering: searchable, citable and updateable enterprise knowledge.
• Agent platform: workflows, tools, model routing, tenancy and observability.
• Model application / algorithm: adaptation, inference, fine-tuning and evaluation.
• AI product: scenarios, metrics, experience, evaluation and business value.
• Quality / evaluation: datasets, scorers, regression and human review.
• Solutions / delivery: customer modeling, integration, deployment and operations.

Capability matrix

• Model: prompts, structured output, context and routing.
• Capability: RAG, tools, memory, MCP and skills.
• Orchestration: state machines, branches, loops, parallelism and approval.
• Engineering: auth, rate limits, timeouts, retries, idempotency and monitoring.
• Business: success rate, efficiency, cost and risk.

Expectations by seniority

• Junior: connect models, build basic knowledge Q&A, wrap tools and produce structured output.
• Mid-level: make the chain stable with state, concurrency, retries, evaluation, cost and permissions.
• Senior: define boundaries between model and code, and design multi-tenant, multi-model, rollout and governance capabilities.

Infer the work from keywords

• LangGraph + state machine + workflow: prepare checkpoints, transitions and recovery.
• RAG + vector database + rerank: prepare Recall@K, MRR, citation accuracy and Badcases.
• MCP + tools + permissions: explain discovery, authentication, isolation and audit.
• Inference engine + quantization + GPU: prepare throughput, memory and degradation strategy.
• Agent evaluation + Trace + quality: show how a non-deterministic system becomes measurable.

Resume and portfolio evidence

Do not write only “built an Agent with LangChain.” Explain the scenario, architecture, action loop, metrics, constraints and result.

Scenario: who had which task
Architecture: model + RAG + tools + orchestration + storage
Actions: how the agent planned, called and verified
Metrics: success, latency, cost and human handoff
Constraints: permissions, sensitive data, failures and rollback
Result: reproducible tests or production evidence

Role-fit checklist

• I can state the business result the role delivers.
• I can draw model, state, tools, data and permission boundaries.
• I have an RAG or tool-calling project with an evaluation method.
• I can explain a failure, not only show a success screenshot.
• I can classify a role as application, platform, algorithm, product or delivery.
• I can describe a project as problem, solution, metrics and retrospective.

Continue with Agent foundations and architecture.