Workflows that other people have actually run, with the commands, the failure modes, and the number of confirmed successes.
Codex is fast and overshoots; Claude is a pedantic reviewer. Chaining them is the exact workflow that saves me hours: Codex writes on a branch, Claude reviews the diff against a fixed checklist, and nothing merges until the checklist passes.
Run three or four agents against the same repo without them stomping each other's changes. One worktree per agent, one branch per worktree, and a cleanup routine that keeps the repo sane.
In a 400k-line workspace, an unscoped agent burns half its context discovering packages it will never touch — then edits one it shouldn't. Run the agent from the package directory, deny writes outside it, and hand it the workspace-level commands it still needs.
Long sessions die one of two deaths: the bill or the compaction. This is the freelancer's version of token discipline — measure per task, keep the context lean on purpose, and checkpoint state to files so a fresh session costs you minutes, not the afternoon.
Agents read everything they can reach, including the .env you forgot about. Layered defense: get secrets out of the tree, deny-list what must stay, and plant a canary so you find out when the layers fail.
Claude Code can't see Xcode's build errors unless you hand them over. XcodeBuild MCP plus a formatter that strips the noise gives the agent a real edit-build-test loop for SwiftUI work — the difference between guessing and iterating.
Agents love 'fixing' installs by regenerating the lockfile, and every regeneration is a 3,000-line diff nobody reviews. Pin the package manager, pin exact versions, and make CI reject any lockfile the humans didn't ask for.
Never let an agent's migration touch a database you care about on the first run. A disposable shadow database catches destructive DDL, broken backfills, and the classic 'rename via drop-and-create' before they reach even your dev data.
Green checks mean nothing if the agent weakened the tests to get them. This gate compares test counts and coverage against the base branch, flags deleted assertions, and refuses to merge on silence-by-subtraction.
MCP servers die quietly mid-session and the agent just... stops using the tool. Run long-lived servers under a supervisor (launchd or systemd), health-check them on a timer, and log restarts so you can see which server is flaky.