Building Software
Engineering Fundamentals for the Agent Era
A catch-all handler that logs the error and returns a default value, so failures look like successes.
Errors re-raised without context, leaving a message like 'failed' with no input, cause or location.
Connections, locks and temporary files left open because cleanup only runs on the success path.
Deciding what happens when things go wrong: which errors to handle, which to pass up, and which should stop the program.
An agent wrapped a payment-settlement function in a handler that catches every error, logs it and returns a status of 'ok', and it opens a database connection before the protected block. Find what callers see when settlement fails and what leaks each time it does.
Watch
Ding Yuan presents the University of Toronto team's study of catastrophic failures in production distributed systems, which traces most of them to incorrect handling of non-fatal errors, including empty or log-and-continue catch blocks. This is the catch-all handler in the drill.
Erlang's creator argues for failing fast and letting a process crash rather than continuing in an unknown state, and explains how errors should be detected, isolated and handled by a supervisor instead of hidden.
Shows how returning Result values instead of throwing makes every failure path visible in function signatures, and how validation and error cases compose without nested try/catch.
The topics 'Dead Programs Tell No Lies', 'Assertive Programming' and 'How to Balance Resources' cover crashing early, asserting impossible states, and releasing every resource on every path.
Case studies of production outages that spread through unhandled failure paths, plus stability patterns (timeouts, circuit breakers, fail fast) that decide what a caller sees when a dependency breaks.
The Defensive Programming chapter separates assertions for impossible conditions from error handling for expected ones, and covers exceptions, barricades at trust boundaries, and robustness versus correctness.
Reference
How to translate internal failures at the boundary into a generic client response with the correct status while logging the detail server-side, so stack traces and internals never reach callers.
Reference
A clear primary explanation of wrapping errors with context (%w) while keeping the cause inspectable with errors.Is and errors.As. The same principle applies in any language that has error chaining.