Public demo of the Grok-to-Codex workflow: sensitive fields are masked.

From spoken intent to live web deliverable

Agentic Trip Harness Demo

How an owner's voice and chat instructions drive research agents, a CTO orchestrator, and subscription coding CLIs to build a verified 16-day Italy itinerary — with personal, timing, and booking details masked.

4-part AI harnessArchitecture · Stack · Output16-day Italy outline5 generic city staysSensitive fields masked

STAGE 1 Architecture design

A four-part harness connects instructions, evidence, tool use, and asynchronous review. People set direction and approve consequential actions; agents gather and organize the supporting work.

flowchart TB
  U["1. User instructions
Owner (Decides) talks to Grok
by voice or chat"] --> K["2. Knowledge (Researches and cites)
Research agents across mail, files, web
Confirmations as source of truth"] K --> T["3. Tool calls
Grok orchestrates · CTO agent (standing rules)
Claude Code & Codex · Booking only with approval"] T --> E["4. Evaluation (Decides)
Travellers review asynchronously"] GH[GitHub
long-term memory] --> K O[Outlook
real-time channel] --> K T --> H["Hosting
Live site: italy-story.richasf.ai"] H -. feedback loop .-> U

1. User instructions

The owner talks to Grok by voice or chat.

2. Knowledge

Research agents work across mail, files, and the web; confirmations serve as the source of truth.

3. Tool calls

Grok orchestrates · CTO agent works under standing rules · Claude Code & Codex act as coding agents · Bookings happen only with traveller approval.

4. Evaluation

Travellers review asynchronously.

People — Decides

Set intent, resolve choices, and approve bookings. Augmented workflow: I decide, the agents do the legwork. Humans in the loop.

Agents — Researches and cites

Gather evidence, cite sources, maintain the plan, and prepare a reviewable deliverable.


STAGE 2 Tech stack

The stack separates research and communication, durable source control, orchestration, coding, and public delivery.

flowchart TB
  subgraph Research[Research and communication]
    G[Grok Bot
voice & chat, research agents] O[Outlook
real-time channel] end subgraph Memory[Source and memory] GH[GitHub
source & CI, long-term memory] end subgraph Control[Orchestration] CTO[CTO agent
standing rules & review] end subgraph Coding[Build tools] CLI[Codex CLI & Claude Code CLI
owner's subscriptions] end subgraph Delivery[Hosting and network] VM[Oracle Cloud server with nginx] CF[Cloudflare
DNS, proxy, TLS] DOMAIN[richasf.ai domain] end ROLE["Decides / Researches and cites"] ROLE --> G G --> O G --> GH GH --> CTO CTO --> CLI CLI --> VM DOMAIN --> CF --> VM

STAGE 3 Output

One public deliverable presents the operating model, a compact trip outline, and a checked redaction boundary.

flowchart TB
  SITE["italy-story.richasf.ai
Decides / Researches and cites"] SITE --> A[Stage 1–3 harness diagrams] SITE --> B[6-step workflow narrative] SITE --> C[Minimal 16-day / 5-stay outline] SITE --> D[Generic masked booking summary] SITE --> E[Zero-leak automated audit gate]

Workflow narrative

The owner decides; the agents research and cite. Review and approval stay with people throughout the augmented workflow.

1

Request

Share intent by voice or chat.

2

Plan

Agents organize the research into a proposed route.

3

Verify

Check evidence against live confirmations and authoritative sources.

4

Book with approval

Proceed only after traveller approval.

5

Review

Travellers review the result asynchronously.

6

Publish

Build and audit the sanitized public page.

Decides: the owner and travellers. Researches and cites: the agents. This is an augmented workflow: I decide, the agents do the legwork, with humans in the loop.

Minimal itinerary outline

A 16-day trip in Italy, 5 city stays (city names, sights, day-by-day details, dates, and traveller identities omitted).

flowchart TB
  A["Arrive
Flight [redacted] · Airport [redacted] · Time [redacted]"] --> B["City stay 1: Capital
Hotel [name hidden]"] B --> C["City stay 2: Southern coast
Hotel [name hidden]"] C --> D["City stay 3: Art city
Hotel [name hidden]"] D --> E["City stay 4: Capital
Hotel [name hidden]"] E --> F["City stay 5: Northern city
Hotel [name hidden]"] F --> G["Depart
Flight [redacted] · Time [redacted]"]

Generic booking verification summary

Status was checked against the live inbox. All identifying values remain masked in this public demo.

Air travel

confirmed via the live inbox (Flight [redacted], Airport [redacted], Time [redacted], Seat [redacted], Cost [redacted])

Lodging

confirmed via the live inbox (5 stays, Hotel [name hidden], Address [hidden], Confirmation [redacted], Cost [redacted])

Intercity rail

confirmed via the live inbox (Train [redacted], Seat [redacted], Time [redacted], Cost [redacted])

Cultural entries

confirmed via the live inbox (Vendor [redacted], Confirmation [redacted], Time [redacted], Cost [redacted])

Traveller identity & timing

Traveller [name hidden], calendar dates masked