Agentic Workflows · ~10 min
Eleven lessons, one reflex: read the situation, reach for the move. This is the whole course as a lookup table — and a mixed review to prove it stuck.
The through-line of the whole course: structure the work before you run it, and put the human where judgment beats persistence. Diagnose the situation, pick the move, and let automation carry the rest. Here's the diagnostic map.
| Situation | Move | From |
|---|---|---|
| Non-trivial change in code the agent may misread | Make it summarize the subsystem read-only, correct it, then plan before any edit | L1 |
| Wrong assumptions are expensive to surface | Research → plan → implement; sandwich reasoning high at plan/verify, low at execute | L2 |
| You keep fixing the same class of bad output by hand | Move onto the loop — fix the harness that produced it, not the artifact | L3 |
| An irreversible vs. a reversible action | Gate by reversibility; review the decision, not the execution | L3 |
| One agent is the bottleneck on independent work | Fan out the factory model — parallel sessions, automated feedback, you orchestrate | L4 |
| Parallel agents would collide on shared files | Worktree isolation — one branch, one PR each; single-branch + guards past ~10 | L5 |
| You want an agent to run on every PR, unattended | Headless in CI — cap --max-turns, shift permissions off prompts | L6 |
| A task's final tenth needs human taste or unreachable infra | Background→foreground handoff at the judgment threshold, with a distilled summary | L7 |
| "Get this unfamiliar repo running" keeps failing silently | Snapshot before state changes; verify with prosecutor-judge, not exit codes | L8 |
| A monolithic agent prototype fails silently in production | The five-step refactor: sequence, schema, dynamic RAG, tracing, circuit breakers | L9 |
| The codebase rots between commits | Entropy reduction agents — encode taste once, scan on a cadence, one-violation PRs | L10 |
| "Done" for an agent feature is undefined or slippery | Write the eval first; baseline before building; the low pass rate is the surface | L11 |
Part 1 was the inner loop you run on one task: plan first, decompose into phases, and decide where you stand relative to the agent. Part 2 scaled that out: parallel sessions move the bottleneck to your review bandwidth, worktrees and single-branch keep swarms from colliding, headless agents bound the pipeline, and handoffs move ownership cleanly. Part 3 kept the codebase healthy across all of it: reversible setup, the prototype-to-production refactor, scheduled entropy reduction, and eval-first development. Each is the same move — structure the work, place the human well — applied at a different layer.
The reflex is to apply the strongest move by default — and every lesson warned against it. Plan-first is overhead on one-sentence changes; research doesn't compound in well-mapped code; the factory model needs deterministic verification or it inflates the 41–87% multi-agent failure band; single-branch is riskier than worktrees without its preconditions; entropy agents can't substitute for fixing a slop factory; evals-first misleads on novel or subjective work. Match the move to the cost of being wrong — heavy structure for durable, expensive work; light touch for throwaway spikes.
Mixed review — across all eleven lessons
Question 1 · from L2The reasoning sandwich spends the most reasoning at…
Question 2 · from L3"On the loop" means that, faced with a bad output, you…
Question 3 · from L6For an unattended CI run, the flag you must set is…
Question 4 · from L8Repo-setup agents verify with prosecutor-judge because…
Question 5 · from L11Writing an eval before the feature means a low baseline pass rate is…