Agentic Systems Engineering
Design typed agents, state machines, coordinated swarms, RAG, repair, and evaluation.
Typed Agent and Tool Design
Build bounded agents whose tools, outputs, memory, and permissions are observable contracts.
01Model the agent loop75 min →02Design strict tool contracts75 min →03Constrain model output70 min →04Bound agent memory75 min →05Enforce permission and cost budgets75 min →Lab: Citation-Grounded Research Agent — Build a research agent that can inspect supplied local documents, invoke typed tools, and refuse unsupported conclusions.
Dynamic State Machines
Make agent behavior recoverable by expressing transitions, checkpoints, and terminal guarantees.
01State machine foundations70 min →02Typed transitions70 min →03Recovery and retry75 min →04Checkpoint durable state75 min →05Branch with bounded policy75 min →Lab: Recovery-Aware Workflow Machine — Implement the lesson workspace state machine with typed events, capped retries, checkpoint restoration, and safe terminal states.
Multi-Agent Coordination
Coordinate specialized agents without uncontrolled delegation, races, or conflicting conclusions.
01Decompose roles by responsibility70 min →02Use explicit delegation protocols75 min →03Protect shared state75 min →04Bound concurrent execution75 min →05Resolve conflicting evidence75 min →Lab: Planner–Researcher–Reviewer Swarm — Coordinate three agents over a shared task ledger with concurrency limits, cancellation, and deterministic merge rules.
Vector Retrieval and RAG
Retrieve relevant evidence, preserve provenance, and measure grounded answer quality.
01Chunk for meaning and retrieval75 min →02Store embeddings in pgvector80 min →03Combine lexical and semantic search80 min →04Rerank and cite evidence80 min →05Evaluate retrieval separately80 min →Lab: Curriculum RAG Assistant — Index supplied course documents in pgvector and answer evaluation questions with exact source citations or an explicit refusal.
Self-Healing Automations
Diagnose controlled change and apply bounded repairs without masking unsafe failures.
01Classify failure before repair70 min →02Repair selector drift75 min →03Repair schema drift75 min →04Bound automated repair75 min →05Escalate with evidence75 min →Lab: Controlled Drift Repair — Repair an academy extractor across supplied page mutations, then prove the repair does not regress prior fixtures.
Agent Evaluation and Guardrails
Measure agent quality, safety, latency, and cost before regressions reach production.
01Trace every decision boundary75 min →02Write deterministic evaluations75 min →03Use model-based evaluation carefully80 min →04Test safety and authorization80 min →05Set cost and latency budgets75 min →Lab: Agent Regression Harness — Create a repeatable evaluation suite spanning tool correctness, grounding, refusal, latency, and cost for the Tier 2 projects.
Evidence-Grounded Agent Swarm
Build a resumable coordinator with specialist agents, typed tools, hybrid retrieval, citations, permission budgets, deterministic model doubles, and regression evals.
10 production starter files · 4 failure drills · automated 100-point rubric
Inspect the capstone →