log/ianHyeonseok JeongEmail

Learning Lab / AI Agents · Workflow Automation · Operations

Hermes Human-Gated AI Agent Workflow

I am rebuilding the Hermes workflow I have operated so I can explain, change, and recover every step myself, not only run it.

  • AI Agents
  • Human Approval
  • State Machine
  • Recovery
Role
Operator and supervised rebuilder
Period
2026 · In progress
Type
Independent AI learning lab
Status
Guided rebuild in progress

Proof at a Glance

Evidence
  • I have operated and checked requirements, approvals, QA, failure states, and delivery verification in the existing workflow.
  • A complete Korean rebuild manual and local reference implementation are ready for my personal practice.
  • My next evidence is a guided rebuild, a requirement change, fault recovery, and a cold rebuild.
Public and Validation Boundary
  • I do not claim that I independently designed the existing production workflow end to end.
  • Credentials, private content, and live publishing connections are not public.
  • The status will change to completed only after I finish and record every rebuild gate myself.

Work flow

Structure and Flow

  1. 01Schedule

    Accept one task for the same date and stage, while blocking duplicate execution.

  2. 02Draft and version

    Create a new version and invalidate old approvals whenever the content changes.

  3. 03QA and approval

    Separate editorial approval from publication approval and bind both to exact hashes.

  4. 04Dry run

    Prepare a package, preview the result, and block publication if the approved version changes.

  5. 05Verify and recover

    Read the result again, compare hashes, record failures, and recover without duplicate delivery.

Why I am rebuilding it

I have used an AI content workflow that connects Hermes, Codex, Discord approvals, and a durable state ledger. Operating it taught me an important lesson: useful automation is not only a model call. It also needs clear states, human decisions, logs, and a way to stop safely.

I can explain the operating requirements and verify the existing workflow, but that is different from building it alone. I do not present the existing implementation as my independent completed work. This lab closes that gap.

What the system must guarantee

The workflow is designed around operational rules, not a flashy demo:

  • The same scheduled task is accepted only once.
  • A changed draft creates a new version and cancels old approvals.
  • Editorial approval and publication approval are separate decisions.
  • Approval is tied to the exact content, package, and dry-run hashes.
  • Publication is blocked when the approved preview changes.
  • Failure states are recorded so the workflow can be recovered without guessing.

What I am doing myself

I am following four evidence gates:

  1. Rebuild the full local workflow while explaining every file and command.
  2. Change one requirement and update the state logic and tests.
  3. Inject failures, diagnose them from logs and state, and recover safely.
  4. Rebuild the system again without the manual.

This page is public before the final gate because it records the work honestly. It is a learning lab in progress, not a completed employment project.

What completion will prove

When the gates are complete, I should be able to set up Hermes, load a trusted project plugin, explain the state machine, change an approval rule, diagnose a broken run, and reproduce the workflow on a clean machine. The final evidence will be commands, test results, state snapshots, and a short recovery log.

Why this matters for AI AX work

Business automation needs more than speed. A person must know what the agent did, approve important actions, detect changed inputs, and recover from failure. This lab turns those ideas into behavior I can test and explain.