Jake Tae-moon Jo · Engineering lead · https://jake-jo-portfolio.vercel.app/en/process/
This is how I set up and ran the work on the 2026 EWA platform, and how that setup changed across six 2026 projects. Every change request from the client was shipped within 24 hours. I defined an 8-stage pipeline and assigned design, implementation, QA and review to a 13-role AI agent team. The client's stakeholders supplied the requirements and the decision-sheet answers; I confirmed or gave input on decisions taken by default, and gave the go for production.
Agents · 8 stages · 13 roles
One feature, from spec to production
The eight stages in the middle are assigned to role agents. At the top the client supplies requirements and I confirm or give input on decisions taken by default; at the bottom I give the go for production.
Click a node, or Tab to it and press Enter, to highlight its note. Rounded, heavy boxes are the client's. Filled rounded boxes are mine. Dotted rounded boxes are rules that never fired.
The rule: every feature traces to a story ID in the spec, and branches, commits and PRs carry the ID. In practice 61.4% of the 1,855 non-merge commits carry a story ID, and 70.8% carry a story ID or a decision number. There were no PRs, and 15 of the 200 branch names contain a story ID.
Instead of blocking, go ahead on a recommended default, log it, and review the defaults later in one pass. 192 of the 199 rows say how hard the decision is to reverse, and 11 are hard, locking in data or contracts. Of those, 3 were confirmed by a person when logged and 8 went ahead on the default. A row that reverses an earlier decision names it.
The rule: write the failing test first. 86.4% of the 398 non-merge fix commits and 87.7% of the 764 feat commits also changed a test file. That includes edits to existing tests, so it does not mean a regression test was added each time.
The one contract between repos is the backend's OpenAPI snapshot. The web generates its types from it; the app's models are written by hand (D-009, code generation deferred). When the contract changes, the API contract check looks for drift between the web, the app and the backend.
Several agent sessions run at once. Each session uses its own worktree and announces the files, decision numbers and migration names it will touch. A decision number is taken as max + 1 at the moment of writing.
An adversarial review of any change that touches withdrawals, settlement, payouts, auth, permissions or personal data; it does not edit code. Since 28 Sep, when records begin, it ran 10 times. 30 fix commits mention security review or a security fix (Jun 14 – Oct 29).
Correctness, repo conventions, simplicity and test quality on every change. Findings are graded blocker, major, minor or nit, and the verdict is APPROVE or REQUEST-CHANGES. It ran 12 times since 28 Sep.
When the contract or a client changes, it checks that the web and the app match the backend's OpenAPI. It ran twice since 28 Sep.
An unresolved blocker or major finding, or any CRITICAL or HIGH security finding, counts as a failed gate and goes back to development. A further rule sends a review-and-fix loop to me once it repeats more than twice on the same finding. Both are rules in the orchestrator's instructions, and no escalation is on record.
Permission settings block force-push, reset --hard and rm -rf, and three of the four repos also block reading .env. A hook asks for confirmation before every delete-like command sent to the staging server.
| Project | Agent roles | What enforced the gates | Last gate, held by | Changed from the row above |
|---|---|---|---|---|
| Apr – Jul 2026Construction site management | 11 · 1 reviewer | Agent instructions (test first · 80% · commit after review and QA) | Me Merged 106 PRs | Start: 7 roles, grown to 11 |
| May – Nov 2026Appliance rental and maintenance | 11 · 1 reviewer (copied) | CI (types · tests · build) · deploy after CI · typed phrase to delete | Me Merged 139 PRsMe Ran the pre-go-live reset | Team copied; added CI on every PR and deploy once it passes |
| Jun 2026 – presentEarned wage access fintech | 13 = 9 + 4 reviewers | Instructions · deny list · delete-command check hook | Me Push · deploy · store go-ahead | Reviewers 1 → 4 · check hook · reversibility per decision |
| Aug – Sep 2026Power market price analysis | 1 analyst agent · 2 review passes | Stops at first failure · 21 gate checks · publish stamp | Me Adopt or revert hypotheses | Rules moved into code checks · one analyst instead of a role team |
| Sep – Oct 2026Screenplay translation | 11 crew roles · a skill conducts | Instructions · tool exit codes · gates tested on broken copies | Me 3 direction callsClient Per-film checks pending | Back to an 11-role team · gates test themselves |
| Oct 2026 (20 days)Social-graph trading | 13, EWA's role names | Instructions · deny list (same as EWA) · one sensitive change at a time | Me Check per bundle · commit on requestClient Answered the planning questions | Deny list back · work in bundles |