Agentic Workflow Manual

Mental Model

Agentic development is delegated work with verification.

The human remains the engineer of record. The agent can explore, draft, code, test, review, and summarize, but the human owns correctness.

ADLC Loop

ADLC loop

IntentHuman owns the outcome.

Context packRelevant files, rules, commands.

Agent taskBounded, measurable assignment.

OutputPatch, map, test, or analysis.

ReviewHuman checks the work.

VerifyTests and runtime proof.

IntegrateAccept or reject.

LearnUpdate docs and next task.

Good Agent Tasks

Good:

Inspect the ticket creation flow. Identify the frontend form, API route,
validation code, database write, tests, and start commands. Do not edit files.
Return a map with file paths and risks.

Bad:

Make ticketing better.

Use Agents For

Mapping an unfamiliar repo.
Finding all call sites of a function.
Drafting a test plan.
Generating first-pass code for a narrow function.
Explaining errors.
Reviewing diffs for edge cases.
Producing runbook drafts.

Do Not Blindly Delegate

Security decisions.
Data deletion.
Production deploys.
Schema migrations.
Pricing or legal claims.
Final merge judgment.
User-facing promises.

Verification Checklist

Before accepting agent output:

Did it touch only expected files?
Can I explain the diff?
Did tests run?
Did the output match the task?
Did it invent APIs or assumptions?
Did it handle errors?
Did it update docs or contracts if behavior changed?
Is there live or local runtime proof?

ADLC Artifact

Create agentic-log.md:

# Agentic Log

## Goal

## Context Given

## Agent Task

## Agent Output

## Human Review Notes

## Verification Commands

## Accepted

## Rejected

## Risks

## Next Task

Done When

The learner can show how the agent accelerated work and exactly how the human verified it.