agentic_run — ticket_4472.diff
Bounded job: fix the failing test in this file, then show the proof.
Agentic Workflows Guide
Agentic coding, described without the hype
Agentic coding means giving an assistant a bounded job, letting it plan and edit, then verifying like a reviewer — resolving in real time on one ticket at a time.
Watch a diff rewrite itself, then learn why. No tool purchase required to use this guide.
method: plan → attach context → run the test → read every hunk → ship with a rollback path
drawn from how Claude Code, Cursor, and Codex are taught in 2026 training
Four stages, one loop you can repeat
Learn AI-powered development as a review habit, not a prompt trick. Each row below is a stage of the same run — the ticket stays small so you can reject it in parts.
Planning: write the job down first
Name the ticket, the files, expected behavior, and how you test it. In our Cambridge classes the plan fits on an index card. If it does not fit, the job is too large for one run.
index card · one screen
reject in parts
Context: show enough, hide the rest
Attach linked modules and the relevant test. Exclude secrets, dumps, and unrelated directories. Learners practice trimming context until the suggestion sharpens.
linked files + one test
less wins
Validation: run, do not trust
Run the test tied to the change and read the output yourself. Ask for candidate causes when it fails, then confirm by executing code. Log what you tried in two lines.
test output read
run-first
Review: read every hunk
Read the diff line by line: what changed, what proves it, what stays risky. Request narrow revisions per file, not a fresh rewrite. Nothing merges with an unread hunk.
per-file revision
pair review
Our AI coding courses and AI coding training teach this loop across a Claude Code course, a Cursor AI course, and an OpenAI Codex course — so when you Learn Claude Code, Learn Cursor AI, or Learn OpenAI Codex, the review standard stays the same.
release_readiness.md
Release readiness for teams
Release readiness adds four checks: secrets absent, dependencies explicit, migrations reviewed by a human, rollback path known. Teams write these into their pull request template during class. That template travels back to work on Monday and fits existing steps without new tooling.
- check — no secrets or customer data in prompts
- check — dependency changes read as proposals
- check — human sign-off on auth and payment paths
Security notes without fear tactics
Keep credentials out of prompts. Treat generated dependency changes as proposals. Require a human on auth and payment paths. These echo secure software development topics taught alongside assistants used in enterprise training. Plain rules, consistently applied, matter more than elaborate decks.
the exchange turns here — the same file, now read by a second pair of eyes
Judgment call
When agentic work is the wrong call
stopped runs count as good judgment
Large ambiguous rewrites, new architecture, and incident response usually need human-led design first. Use the assistant for mapping, summarizing, and small scoped probes instead. Our instructors mark stopped runs as good judgment, not failure. Knowing when not to delegate is part of the methodology.
Try the loop once before any course.
Pick one small ticket, write a card-sized plan, run one bounded agentic pass, and review every hunk. Note where it helped and where you intervened. You will get more from any track once you have felt the loop once.
Bring those notes to class or to a syllabus call at +1-617-253-1000, 77 Massachusetts Ave, Cambridge, MA 02139. Failing test on entry, passing diff on exit — that delta is the whole method.
failing run → passing run · one session
plan fits one card · context is one test