Walk-through

Demo + presentation script

A guided product tour of Hermes Loop plus a ready-to-read script for 60-second, 3-minute, and 7-minute demos. Seed the workspace first so every page has live data.

See the system map →Operator checklist →Audit /demo-target →

Tour outline (9 steps)

1 · Inbox · messy request lands
Operators drop raw requests here. The desk persists each one as an InboxItem and queues a triage job.
Open inbox →
Tip — once you click Start guided demo, a small card pins itself bottom-right and walks you through every step. Dismiss with ✕ or restart with ↻.
2 · Triage · request becomes a structured mission
The Triage Agent classifies the request, proposes a crew + objective + cadence, and surfaces missing info. You approve, edit, or reject.
See the triage proposal →
3 · Jobs · persistent background queue
Triage, mission runs, and schedule fires all flow through a retryable job queue. Run jobs manually with the Run-due button.
Open job queue →
4 · Mission · agents execute step by step
Each crew member runs in sequence with strict Zod-validated outputs. Replay the timeline to audit any past run.
Pick a mission from the dashboard →
5 · Tools · safe, sandboxed, logged
Agents request tools by name. The desk validates input, runs the tool, persists I/O, and gates risky tools behind approval.
Open the tool sandbox →
6 · Approvals · risky outputs wait for you
Drafts, trade tickets, exports, follow-ups, and gated tool calls land here. Nothing leaves the desk on its own.
Open the approval queue →
7 · Memory · governed, not whispered
Operator preferences become MemoryItems via an explicit suggestion → accept loop. Diff, hygiene, and Hermes sync are first-class.
Inspect the memory vault →
8 · Receipts · proof of every workflow
Every mission gets a sha256-hashed receipt: agents, tools, approvals, memory used, deliverables, errors, risk reasoning. Export Markdown or JSON.
Browse workflow receipts →
9 · Trust ledger · reliability across the system
Aggregated risk, approval rates, tool usage, memory churn, hygiene flags, Hermes sync. The whole desk on one page.
Open the trust ledger →

Presentation scripts

What to say while you click

60 seconds

stops · /dashboard → /missions/[any] → /receipts/[id]

01Hermes Loop is a control room for trustworthy persistent agents.
02Hermes gives agents autonomy. We give them control, governance, approvals, and proof.
03Every mission produces a hashed Workflow Receipt with the full timeline.
04Risky actions wait for you. Memory is governed, not whispered. Reliability rolls up to a Trust Ledger.

3-minute demo

stops · /dashboard → /inbox → /missions/[any] → /approvals → /receipts/[id]

01Land on the dashboard — point at Attention Required, Active Operations, Trust & Proof.
02Open Inbox — show the Triaged proposal with crew, objective, missing info.
03Open a Mission — point at agent timeline, tool calls, memory used (with origin tags).
04Open Approvals — show the gated draft + the inline 'memory suggestion' that the user accepts.
05Open the Workflow Receipt — copy the integrity hash, mention export Markdown/JSON.

7-minute deep demo

stops · /system → /inbox → /jobs → /missions/[any] (Replay) → /hermes/tools → /memory + /memory/hygiene + /memory/diff → /approvals → /receipts/[id] (Risk reasoning) → /trust

01Open /system — the whole product on one page with live counts. Use this to set the mental model.
02Drop a request on Inbox. Show the queued TRIAGE_INBOX job appearing on /jobs.
03Open a Mission. Click Replay — the timeline animates agents + tools step-by-step.
04Open /hermes/tools — point out approval-gated tools are blocked in the sandbox.
05Open /memory — show suggestions, hygiene (duplicate / conflict / stale), then /memory/diff.
06Open /approvals — accept a draft. Show the post-approval memory suggestion appearing.
07Open the Workflow Receipt — explain risk reasoning factors + show export Markdown.
08Open /trust — wrap up: 'this is what a control plane for persistent agents looks like.'

Hermes integration

·Adapter calls Hermes' OpenAI-compatible /chat/completions endpoint.
·Demo mode is a drop-in fallback — same interface, same Zod-validated outputs.
·Hermes Skill Marketplace + Memory Inspector let you govern Hermes-side artifacts from one place.
·Sync from Hermes is feature-detected — UI degrades to local-only if /memory isn't exposed.
·All runtime requests stream into the in-process log on /hermes for observability.

Why this isn't 'just Hermes Agent'

·Hermes does autonomy. Hermes Loop adds the operational frame around it.
·Persistent inbox + job queue → recoverable across restarts.
·Approvals + hygiene + receipts → audit trail you can show a stakeholder.
·Memory governance with diff and approval queue → reliability across repeated missions.
·Trust ledger aggregates risk, approval rate, and tool reliability across every crew.

Safety guarantees

·No live trading — paper trades only, marked simulation only.
·No outbound emails or form submissions — everything routes through the approval queue.
·web_snapshot blocks localhost, private IPs, file://, metadata endpoints.
·Approval-gated tools refuse to run in the sandbox.
·Every prompt + raw response + parsed output is persisted for audit.

Known limitations (be honest)

·Job runner is single-process and triggered manually or via Run-due.
·Hermes /memory and /skills endpoints are stubs — UI handles gracefully if absent.
·Receipt finalization is one-way; no version history yet.
·No multi-tenant auth — single operator profile in this MVP.
·Browser automation is intentionally absent. web_snapshot is read-only.

Demo data included in the workspace

· 3 demo missions (Paper Trading, Bug Hunter, Life Admin)
· 1 active scheduled mission
· 2 inbox items (NEW + TRIAGED) + 2 mobile commands
· 3 background jobs (queued, completed, scheduled)
· 7 memories (LOCAL + HERMES origin, hygiene examples)
· 3 PENDING memory suggestions
· 3 workflow receipts (1 FINAL + MEDIUM risk · 2 DRAFT + LOW)
· 8 marketplace skills + 4 system skills
· 3 sandbox tool runs (one BLOCKED to demo the safety gate)