Dependable Agent
Engineering
Build one bounded agent system from loop to grounded context, safe tools, regression gates, and staged release.
Start chapter 1 →Intermediate · 5 chapters · about 6–8 hours plus project work · Learn at your own paceTHE LOOP
01Implement a bounded agent loop with explicit state, tools, budgets, and stop conditions.
02Construct context and tool contracts that preserve provenance, scope, and authority.
03Build a regression set that measures requirements instead of demo quality.
04Make a staged release decision backed by canary and rollback evidence.
One system,
built in dependency order.
Each chapter includes a worked incident, formative checks, prediction-first break lab, independent build brief, and structured evidence receipt.
01Agent Engineering: From Prompt to Reliable Work Loop40 min guided + independent project+
Watch an unbounded agent fail, put it under contract, break it on purpose, and ship a bounded work loop of your own — with a receipt that proves what it did.
- OUTCOMETrace an agent run and locate the first contract failure.
- OUTCOMESpecify goal, tool, budget, and stop-rule boundaries.
- OUTCOMEImplement and test a bounded loop that fails closed.
02Context, Memory, and Retrieval for Agents40 min guided + independent project+
Watch an agent trust a stale retrieved document as fact, put retrieval under a provenance-and-freshness contract, break it on purpose, and ship a context pipeline that fails closed instead of inventing.
- OUTCOMEDistinguish retrieved text from verified evidence.
- OUTCOMEApply scope, provenance, freshness, and conflict gates.
- OUTCOMEBuild a context packet that preserves uncertainty and citations.
03Tools and MCP: Design the Integration Contract40 min guided + independent project+
Turn a tool from a loose function call into a narrow, checkable contract — schema, scope, idempotency, and typed errors — and prove it holds under three deliberate attacks.
- OUTCOMETurn a function into a narrow tool contract.
- OUTCOMEEnforce schema, scope, permission, idempotency, and typed errors.
- OUTCOMEProve a consequential tool resists duplicate and over-broad requests.
04Evals and Reliability: Turn an Agent Requirement into a Gate40 min guided + independent project+
Turn a fuzzy requirement into a scored regression set with a baseline and a gate, then prove it catches the exact failure a five-prompt demo would have missed.
- OUTCOMERewrite a fuzzy requirement as an executable rubric.
- OUTCOMEConstruct a representative regression set with negative cases.
- OUTCOMECompare a baseline and candidate using an explicit release gate.
05Safe AI Shipping: Build the Gate Before the Launch40 min guided + independent project+
Turn a model or prompt change into a staged, monitored release with a rollback path you've actually tested — instead of a Friday deploy you can't take back.
- OUTCOMEMap model-change failure modes before deployment.
- OUTCOMEDesign a canary with observable promote and rollback criteria.
- OUTCOMERehearse rollback and report what the rehearsal does not prove.
Practise, inspect, revise.
Formative checks distinguish independent answers from answers completed with support. Break labs make you predict before revealing the worked result. Project receipts help you inspect the evidence you produced without pretending the Academy executed your code. The point is better judgment and better work, not a badge.