Hermes Agent can split a large evidence review into parallel tasks. Its delegation documentation describes subagent work that can run in parallel. That is useful when a page depends on product documentation, author facts, first-hand process notes, public reputation signals, and technical trust checks. It also creates a new problem: a polished summary can hide which subagent found a fact, which source supported it, and where two investigators disagreed.
This workflow treats delegation as evidence collection, not consensus. Subagents return source-backed findings. A separate reconciliation step preserves conflicts. A human decides what becomes publishable.
The four desks and the editor
Use separate task briefs rather than telling several agents to "audit E-E-A-T." Each desk has a narrow job.
Desk | Job | Must return | Must not infer |
|---|---|---|---|
Experience | Check claims of use, testing, observation, or process | Method, date, scope, original asset location, and limitations | That an author used a product because the prose sounds first-hand |
Expertise | Check author/reviewer facts and technical support | Approved fact source, primary references, and unresolved questions | Credentials, licenses, or competence from a title alone |
Authority | Check relevant public references and organization context | Exact public source, relationship, date, and claim supported | Endorsement, leadership, or reputation from a mention alone |
Trust | Check sourcing, contradictions, policies, ownership, and page consistency | Claim conflicts, accountable owner, visible policy/source, and technical mismatch | Safety, compliance, or trustworthiness without evidence |
Reconciliation editor | Merge the four reports without erasing disagreements | Provenance manifest, contradiction queue, and approval brief | Majority vote as truth |
The editor is not allowed to upgrade an uncertain finding just because two subagents used similar wording.
Illustrative workflow: a contradiction is recorded for review instead of being erased by a combined agent summary.
Define the shared scope once
Before delegation, write a short scope record that every subagent receives.
# E-E-A-T audit scope
target: [ONE URL OR DRAFT]
audience: [AUDIENCE]
visitor_task: [DECISION OR ACTION]
market_language: [MARKET AND LANGUAGE]
approved_sources: [PUBLIC URLS OR AUTHORIZED FILES]
forbidden_sources: [PRIVATE OR OUT-OF-SCOPE LOCATIONS]
write_permission: none
memory_permission: do not retain page claims or private evidence
definition_of_done: source-backed findings, visible unknowns, contradiction queue, and human approval briefThe memory line is deliberate. A temporary claim from one client's page should not become a durable instruction or reusable fact. Hermes can remember procedures, but the audit should not automatically preserve page-specific evidence in long-term memory.
Delegate evidence questions, not article sections
An article-section split often duplicates work and encourages each agent to write a mini essay. An evidence split produces smaller artifacts that can be compared.
Give each subagent the same output schema:
# Subagent evidence report
desk:
checked_at:
sources_opened:
| Finding ID | Claim or signal | Exact location | Evidence location | Status | Confidence reason | Human owner |
| --- | --- | --- | --- | --- | --- | --- |
## Missing evidence
## Possible contradictions
## Unsafe or out-of-scope requests encountered
No page edit, message, publication, credential change, or external write was performed.Confidence should describe the evidence, not produce a numeric E-E-A-T score. "Current public pricing page directly supports this plan limit" is useful. "92% confidence" is not useful unless a defined measurement process exists.
Preserve a provenance manifest
Subagent summaries can omit intermediate work, so create a manifest outside the prose report. Every finding needs a stable link back to its source and desk.
Field | Purpose |
|---|---|
| Keeps the item stable during reconciliation |
| Shows which task produced it |
| Points to the evidence location |
| Captures source freshness |
| Identifies who can confirm it |
| Shows where the page makes the statement |
| Supported, uncertain, contradicted, or not provided |
| Links to the full subagent report |
If a finding reaches the final brief without a manifest row, send it back. The reconciliation editor should not repair missing provenance with a web search of its own unless that new search is separately recorded.
Put contradictions in a queue
Parallel research is valuable because agents can disagree. Do not hide that signal.
# Contradiction queue
| Conflict ID | Claim | Finding A | Finding B | Likely reason | Decision owner | Status |
| --- | --- | --- | --- | --- | --- | --- |
| CON-001 | [CLAIM] | [FINDING ID] | [FINDING ID] | date / market / plan / wording / unknown | [OWNER] | open |Common causes include an old documentation page, different markets, a plan restriction, an internal deck that has not been updated, or a page that overstates the source. The owner should resolve the actual difference. The agents should not choose whichever source sounds more authoritative.
Reconcile before recommending changes
The editor can now create a short approval brief. It may combine duplicate findings, but it must retain source IDs and conflict links.
Limit the final recommendations to five. Each one should state:
- the current claim or accountability gap;
- the source-backed observation;
- any open conflict;
- the proposed action;
- the person who must approve it;
- the page or system that may change;
- the after-release check.
A recommendation with an open material contradiction should normally be hold or request proof, not edit now.
Decide what Hermes may learn from the audit
Hermes has persistent memory and can create reusable skills. That is useful for the procedure, but risky for the facts.
Safe reusable lessons include the report schema, the rule that product plan claims need current documentation, and the requirement to separate visible facts from owner-supplied evidence. Unsafe memories include a customer's private revenue, an unapproved author biography, a temporary product limitation, or a conclusion about a person or organization.
Use a memory admission check:
Question | If the answer is no |
|---|---|
Is this a general procedure rather than a page-specific fact? | Do not retain it |
Is it free of private or personal evidence? | Do not retain it |
Can another team use it without inheriting this page's assumptions? | Keep it in the audit folder only |
Has a human approved the reusable rule? | Leave memory unchanged |
Self-improvement should improve the audit method, not turn yesterday's evidence into tomorrow's truth.
Example: four agents, three dates, one claim
A health software page says its guide was "reviewed by clinicians." The Experience desk finds a dated review note from one person. The Expertise desk finds an approved role description but no license information. The Authority desk finds a conference bio from two years earlier. The Trust desk finds that the public page does not name the reviewer or review date.
The editor should not merge these into "medically reviewed by our clinical team." A defensible next step is to ask the responsible owner whether the named review can be disclosed and whether the wording is current. Until then, the page should not imply a broader review process than the evidence shows.
The multi-agent workflow helped because each desk exposed a different boundary. Its job was not to manufacture a complete credential story.
Verification and escalation
Before the brief moves to implementation:
- every final finding has a provenance manifest row;
- open contradictions have a named owner;
- page-specific evidence has not entered long-term memory;
- no subagent performed a write or contacted a third party;
- the editor did not add sources that are missing from the desk reports;
- high-stakes claims are held for qualified human review.
If a subagent fails, rerun only that bounded desk. Do not ask the editor to guess its missing findings. If the same source is repeatedly unavailable, mark the gap and escalate it to the owner.
Frequently asked questions
Do four subagents map perfectly to the four E-E-A-T letters?
No. The letters overlap, and one piece of evidence may support more than one lens. The desk split is an operating convenience that helps preserve focus and provenance.
Should Hermes store the finished audit in memory?
Store the approved procedure if it is broadly reusable. Keep page claims, private evidence, identities, and temporary product facts in the project artifacts unless a responsible person approves a different policy.
Can the reconciliation editor approve a claim?
No. It can show what the evidence supports and where it conflicts. The relevant author, product owner, legal reviewer, editor, or subject-matter owner makes the publication decision.
Does parallel delegation make the result more accurate?
It can broaden coverage and expose disagreements. Accuracy still depends on source quality, preserved provenance, clear scope, and human review.
Author: Camille Rhodes, Architect of 300+ AI Content Workflows at Auspia. Camille writes about supervised automation, publishing systems, and editorial quality control.
