A 6-month roadmap to becoming an agent engineer, including Jev routing

Suraj Sharma @suraj_sharma14Originally posted 2026.10.12

Open original on X

If I had 6 months to become an AI Engineer in the Agentic Era. I’d do this. Stage 1: Context Engineering & Token Economics - Learn: Token math, context window limits, dynamic assembly, context compression, budget allocation. - Practice: Build a dynamic context assembler that strictly budgets tokens across working memory, tool schemas and retrieved docs per request. - Why: Prompt engineering is dead. Context overflow is the #1 silent killer of production agents. Stage 2: Durable Agent Orchestration - Learn: State machines, Temporal, LangGraph, checkpointing, human-in-the-loop (HITL) gates. - Practice: Build a 4-hour agent workflow that survives server crashes, API timeouts and waits 3 days for human approval before resuming. - Why: Endless `while True` chat loops break in production. Real business workflows require durable execution. Stage 3: Tool Use and MCP (Model Context Protocol) - Learn: JSON-RPC, strict Pydantic schemas, tool discovery, scoped credentials, MCP servers/clients. - Practice: Write a raw MCP server from scratch exposing your database and internal APIs without using a wrapper SDK. - Why: Agents need standardized, secure hands to do real work. Bespoke API wrappers don't scale. Stage 4: System-One vs. System-Two Routing - Learn: Typed decision models (Jev-style), confidence scoring, latency vs. cost tradeoffs, model cascades. - Practice: Build a triage router that sends 95% of traffic to a fast System-One model and escalates only the 5% edge cases to a heavy reasoning model. - Why: Sending every query to a frontier reasoning model burns your runway in a week. Unit economics dictate survival. Stage 5: Advanced Agentic RAG - Learn: Contextual retrieval, GraphRAG, query rewriting, self-correction, hybrid search, abstention logic. - Practice: Build a RAG agent that critiques its own search results, rewrites ambiguous queries and confidently refuses to answer if grounded context is weak. - Why: Naive chunk-and-retrieve RAG is a liability. Agents must know when to search again. Stage 6: 3-Tier Memory Architectures - Learn: Working memory (scratchpad), Episodic (compressed session logs), Semantic (vector DB), eviction policies. - Practice: Build a memory engine that automatically compresses, scores and evicts old context to keep the agent sharp over long sessions. - Why: Agents without memory fail by turn five. Agents with infinite memory hallucinate from context bloat. Stage 7: Trajectory Evals & CI Gates - Learn: Step-level grading, LLM-as-a-judge calibration, golden datasets, regression testing, shadow routing. - Practice: Build a CI pipeline that grades the agent's tool-call sequence and blocks PRs if it deviates from the golden path. - Why: Final-answer evals hide the exact step where the agent went rogue. The trajectory is the unit of accountability. Stage 8: Inter-Agent Security & Guardrails - Learn: Prompt injection defense, PII masking, least-privilege tool scoping, sanitization proxies. - Practice: Build a middleware that swaps PII for UUIDs before the LLM sees it and prevents Agent A's output from injecting Agent B. - Why: Multi-agent swarms create internal attack surfaces that standard unit tests completely miss. Stage 9: Real-Time Multimodal & Voice UX - Learn: Server-Sent Events (SSE), WebSockets, Voice Activity Detection (VAD), interruption handling, optimistic UI. - Practice: Build a voice agent with sub-1s interruption handling that streams tool calls and text simultaneously. - Why: The interface is moving off the screen. Perceived latency and interruption handling are the new UX standards. Stage 10: The DPO Data Flywheel - Learn: Preference pairs, LoRA fine-tuning, automated feedback loops, synthetic edge-case generation. - Practice: Automate a pipeline that captures user thumbs-downs, formats them into preference pairs, and triggers a nightly model fine-tune. - Why: Your production edge cases and failures are the highest-leverage training data you own. Stage 11: FinOps & Graceful Degradation - Learn: Semantic caching, cost kill-switches, fallback chains (Frontier → Mid → Local → Cache). - Practice: Build a middleware that halts an agent loop instantly if projected spend exceeds $0.50 and degrades to a cached response. - Why: A runaway agent loop is an outage with an invoice. Stage 12: Public Teardowns & System Portfolios - Learn: Architecture decision records (ADRs), benchmarking, system diagramming, public writing. - Practice: Publish 3 deep-dives on how you reduced agent latency, cut inference costs or improved tool-call accuracy by 80%. - Why: Senior AI engineers are hired for their system design judgment not their API wrapper skills. Wrappers are a commodity. Systems engineering is the moat. The modern AI Engineer builds the deterministic harness around the stochastic brain. Bookmark and Repost!