Agentic Engineering Beyond Vibe Coding Methodology 2026

Purpose

This PoC tested the smallest useful version of an "agentic engineering" control loop: a proposed code change is only accepted after deterministic verification, and the verification trace is saved as an auditable artifact.

The sandbox did not call an AI model. It simulated a good baseline and a bad agent patch so the article can discuss the engineering boundary without pretending that Effloow ran a production agent workflow.

Commands

rm -rf /tmp/effloow-agentic-engineering-poc
mkdir -p /tmp/effloow-agentic-engineering-poc
cd /tmp/effloow-agentic-engineering-poc

Created:

• package.json
• calculator.mjs
• test-runner.mjs
• gate.mjs

Baseline verification:

npm run gate

Output:

> gate
> node gate.mjs

GATE_PASS deterministic verification accepted release

Injected a deliberately bad patch:

perl -0pi -e 's/Math\.round\(cents - \(cents \* percent \/ 100\)\)/cents - percent/' calculator.mjs
npm run gate || true

Output:

> gate
> node gate.mjs

GATE_FAIL deterministic verification blocked release

Trace captured by gate.mjs:

[
  {
    "label": "unit-test-gate",
    "cmd": "node test-runner.mjs",
    "status": 1,
    "stdout": "PASS zero percent",
    "stderr": "FAIL ten percent: expected 900, got 990\nFAIL round half up: expected 874, got 986.5"
  }
]

Restored the correct implementation:

perl -0pi -e 's/cents - percent/Math.round(cents - (cents * percent \/ 100))/' calculator.mjs
npm run gate

Output:

> gate
> node gate.mjs

GATE_PASS deterministic verification accepted release

What Worked

• A minimal deterministic gate separated a plausible-looking patch from an acceptable patch.
• The gate produced a machine-readable trace.json artifact containing command, status, stdout, and stderr.
• The failing patch was blocked before any release step.

What Failed

• The trace was overwritten on each gate run instead of being appended with immutable run IDs.
• The sandbox used one unit-test gate only. It did not include linting, static analysis, security checks, snapshot tests, or human review.
• The bad patch was manually injected rather than generated by a real coding agent.

Limitations

This PoC proves a workflow shape, not a product benchmark. It does not measure agent coding quality, speed, cost, reliability, or defect rate. It also does not prove that a specific commercial coding agent follows this process. The article should only claim that Effloow Lab ran a local sandbox gate demonstrating why professional agentic workflows need deterministic verification and trace capture.