How-to

Human in the loop

Keeping a person between the agent and a customer, a published output or an irreversible action.

Type
How-to
Confidence
High
Sources
7 sources
First published
Updated

Across eight posts the dominant pattern is that the agent drafts, researches or coaches and a person decides. The Halfbit Studio author says the agent does not send emails and client contact stays human. The physician author keeps responsibility for accuracy and final recommendations, with AI producing first drafts, and warns that polished, fast output makes oversight more important. The attorney running a subscriber Q&A uses a human tag gate: he reviews rows before a separate cron publishes them. Lead vetting findings go to a human for validation, and a support chain escalates from a local-model agent to Hermes to the team. Craig Hewitt shows the gap in practice: his first ticket scan produced eight suggestions, but the process for validating and acting on them is not yet decided. Botsitting, long chat sessions fixing errors, is the cost side. Evidence is self-reported and single-case.

Key points

  • Keep customer contact and final decisions human; the agent drafts or advises — 4 posts. (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L141) (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L120) (an attorney who runs a local business · My Hermes setup (Notion knowledge base)… ¶ L79) +1 more
  • Put an explicit review gate between agent drafts and anything published — 2 posts. (an attorney who runs a local business · My Hermes setup (Notion knowledge base)… ¶ L80) (an SME automation consultant · Anyone else deploying Hermes into SME… ¶ L14)
  • Escalation chains pass hard cases up to people — 1 post. (a founder who replaced their helpdesk · We replaced our helpdesk with Hermes… ¶ L6)
  • Oversight matters more as output gets faster and more polished — 1 post. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L192)
  • Validation processes are often undecided or costly (botsitting) — 2 posts. (Craig Hewitt · Day 4 of agentifying my business ¶ L10) (Marc Bara · Marc Bara: 15 days with Hermes Agent, 50+… ¶ L124)

Where posters agree

  • Where the agent touches customers or published content, posters keep a person in charge of the final step (author of the Halfbit Studio blog post · Halfbit Studio: How we run our company with… ¶ L141) (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L120) (an attorney who runs a local business · My Hermes setup (Notion knowledge base)… ¶ L79) +1 more

Frameworks

  • Private-to-public Q&A pipeline with human tag gate — Hermes backfills and nightly-updates a private Q&A database from sent email; the poster reviews and tags rows, and a separate cron sweeps tagged rows into the public subscriber-facing database that the chatbot reads. (an attorney who runs a local business · My Hermes setup (Notion knowledge base)… ¶ L80)
    • Private Notion database: question, answer, category columns
    • Hermes backfills from past email threads and checks the sent folder each night
    • Hermes leaves category blank; the poster reviews, revises and tags
    • A cron sweeps tagged Q&A into the public-facing database
  • Hermes-fronted support escalation chain — Hermes gathers the customer’s details and starts a per-customer Pi session on a local model; that agent can escalate to Hermes, which can escalate to the humans. (a founder who replaced their helpdesk · We replaced our helpdesk with Hermes… ¶ L6)
    • Customer messages the company
    • Hermes gathers relevant customer details
    • Pi session on local Gemma 4 talks to the customer
    • Pi agent escalates to Hermes
    • Hermes escalates to the team

Stories & examples

  • Lead vetting for a drinks packaging company: Prospects from food expos are checked against their websites, findings go to a human for validation with reasoning and an image, approved contacts are enriched through Lusha and qualified leads are pushed into the CRM. Lesson: A human validation step sits between research and CRM entry. (an SME automation consultant · Anyone else deploying Hermes into SME… ¶ L14)
  • First automated support-ticket scan: The first scan produced 6 help-doc suggestions and 2 marketing suggestions; the open problem is the process for validating and acting on them. Lesson: The author sees the validation and action process as the next challenge. (Craig Hewitt · Day 4 of agentifying my business ¶ L10)
  • Interactive botsitting: A commenter ends up in long interactive chat sessions fixing the agent’s errors; Marc Bara says he has had the same feeling of botsitting while models work. Lesson: He now runs four or five projects at once. (Marc Bara · Marc Bara: 15 days with Hermes Agent, 50+… ¶ L124)

Caveats & warnings

  • The process for validating and acting on the suggestions is not yet decided. (Craig Hewitt · Day 4 of agentifying my business ¶ L10)
  • The author frames the use as not autonomous medical decision-making and not the practice of medicine. (Dr. Chukwuma Onyeije · Dr. Chukwuma Onyeije: I gave an AI agent 94… ¶ L106)

How much to trust this

  • Evidence: 8 posts, 9 independent speakers (repeat posters counted once).

Blueprints that use this

Sources

  1. (01)

    Halfbit Studio: How we run our company with Hermes Agent

    Post

  2. (02)

    Anyone else deploying Hermes into SME businesses?

    Post

  3. (03)

    Dr. Chukwuma Onyeije: I gave an AI agent 94 skills

    Post

  4. (04)

    My Hermes setup (Notion knowledge base), crons and use cases: attorney-operator on a Raspberry Pi

    Post

  5. (05)

    Day 4 of agentifying my business

    Post

  6. (06)

    Marc Bara: 15 days with Hermes Agent, 50+ capabilities

    Post

  7. (07)

    We replaced our helpdesk with Hermes (comment on 'Salesforce to acquire Fin')

    Post

(By email)

Compounding Insights