Field guide · applied agent engineering
A working reference for the full course—from prompting through production security. Use the patterns while you build; use the checks before you ship.
Mira can communicate with a defined role, task, constraints, and output shape.
Understand probabilistic output, tokens, context limits, and the five-part prompt structure used throughout the course.
Compare prompt structures and convert an unstructured response into a validated result.
Mira can choose a permitted tool and use its result without seeing its credentials.
Move from a single model response to a controlled loop that can select tools, act, observe results, and stop safely.
Design a minimal tool set, schemas, permission boundaries, and failure behavior.
Read the complete guide ↗Mira can break a goal into bounded steps and explain what happened.
Compose routing, chaining, decomposition, and evaluator patterns into a small agent that is easy to inspect.
Build the travel-agent core, then test tool errors, incomplete inputs, and unsupported requests.
Mira can remember an approved preference without leaking it into another user’s session.
Separate working context, session state, durable preferences, and source-of-truth data instead of calling everything memory.
Add explicit save/retrieve behavior and verify isolation across sessions and users.
Read the complete guide ↗Mira can answer from approved documents, cite evidence, and abstain when evidence is weak.
Build ingestion, chunking, embeddings, retrieval, grounding, and citations as one measurable system.
Create a small knowledge base, inspect retrieved chunks, and test unanswerable questions.
Read the complete guide ↗Mira can operate within a latency target, cost budget, and defined fallback path.
Treat cost, latency, reliability, observability, and deployment as architecture—not cleanup after the demo.
Measure a baseline, remove unnecessary context, add caching, and compare before/after behavior.
Mira can refuse unsafe actions, protect secrets, and surface uncertain decisions to a person.
Test task quality and system safety with datasets, thresholds, permissions, adversarial cases, and human escalation.
Create quality, injection, exfiltration, excessive-agency, and escalation tests.
Mira can complete a bounded workflow and show the evidence needed to review it.
Combine goals, tools, memory, knowledge, guardrails, evaluation, and operations into one defensible system.
Deliver an architecture brief, working flow, evaluation results, threat review, and operating plan.
The final test
A production-ready agent is not the one with the most components. It is the smallest system that meets the goal, respects its boundaries, and leaves enough evidence to operate safely.
Join the course list