Language: English · 中文
This page walks through one working day on a multi-agent stack—how humans and agents stay aligned without turning the shared room into a build log. It is a pipeline story, not a product pitch. Architecture details live on the public spine page.
Intended public home (after content-gated publish): https://machengshen.github.io/pipeline/ · Architecture spine: https://machengshen.github.io/spine/
You do not need to run agents to understand this. Think of it as a collaboration pattern with seven stations. Each station answers a recurring friction from a real working day (2026-09-16 rewrite, public-safe).
Station 1 · Multi-agent group chat as standup
Morning starts in a shared room with several agents and the owner. The room’s job is align only: priorities, go/hold, who owns which lane. Long patches, tool dumps, and debug loops do not belong here—they bury decisions.
Falsifiable check: if a random day of the shared thread is mostly build noise, the standup claim has failed.
Station 2 · Shared Memory Hub for standing prefs
Chat is ephemeral. Preferences that must survive the next session go into a Shared Memory Hub: durable standing decisions, task receipts, and a shared profile other agents can query.
Owner phrase pattern (Chinese): when someone says they have “noted it” (「记下了」), the intended meaning is a triple write—not a polite chat acknowledgment:
If step 1–3 never happen, “noted” is a failure mode—even if the chat looked polite.
Station 3 · Outer harness when the core product harness is closed
Some days the primary product harness (the usual IDE / cloud coding surface) is unavailable, rate-limited, or waiting on auth. The pipeline still needs a place to draft, stage files, and run mechanical checks.
Outer harness here means a secondary shell—another agent runtime or scratch workspace that can hold work until the core harness reopens. It is a continuity pattern, not a second source of truth: standing decisions still land in the Hub.
Station 4 · Public spine page for architecture
When the question is “how is the stack shaped?”, do not re-explain in Slack. Point to the public architecture page:
https://machengshen.github.io/spine/
That page covers shared memory, standup vs private lanes, typed packets, and take/leave guidance with falsifiers. This pipeline page is the day-shaped walkthrough; the spine is the map.
Station 5 · Slack + Cursor Cloud Agents need a repository
Cloud coding agents attached to Slack are repo-shaped, not free-floating REPL windows. Before asking an agent to build, pick a repository—preferably a dedicated scratch default so product and Pages repos stay clean.
Station 6 · Phone-first delivery
When humans will open a link on a phone, verify the URL returns HTTP 200 before sending it. Broken or auth-walled links waste attention and erode trust. Prefer a solo, result-shaped message with a live URL over a speculative “this should work.”
Station 7 · Content-gated publish (PCSE-style)
Public pages ship only after a mechanical clean-content checklist passes. Public description of the gate:
Compared with single-thread chat products (one conversation, one model, memory mostly inside that thread), a multi-agent group plus connector-style Hub UX is different in ways you can test:
These are product differences you can falsify in your own stack. This page does not claim absolute uniqueness.
Machine-oriented companion to the human narrative. Restate, then execute the checklist. Do not invent private infrastructure from this page.
| Pitfall | Symptom | Fix pattern |
|---|---|---|
| Standup becomes a build log | Patches / tool dumps bury go-hold decisions | Move builds to private lanes; return a short receipt to Hub + standup |
| “Noted” without writes | Preference vanishes next session | Hub standing + shared profile + sibling notify (triple write) |
| Core harness closed → freeze | No drafts or checks land anywhere | Use an outer harness shell; keep Hub as system of record |
| Re-explain architecture in chat | Repeated long essays in Slack/group | Link /spine/; keep chat for decisions |
| Treat Cursor Slack as a REPL | No repo; agent cannot attach work | Pick a dedicated scratch default repository |
| Personal Cursor login only | Pick Repository fails despite correct URL | Install Cursor GitHub App on the GitHub org; reconnect if needed |
| Send URLs before live | Humans hit 404 / auth walls on phone | Verify HTTP 200, then send a solo result message |
| Ship without clean gate | Secrets, private hosts, or absolute-uniqueness claims leak | Run PCSE-style checklist; fail closed |
| Gate | Pass condition |
|---|---|
| clean_content | Forbidden-token scan empty on public HTML |
| secrets_scan | No API keys, tokens, passwords, private URLs |
| public_surface_only | Links are public github.io / public GitHub only |
| no_absolute_uniqueness | Claims are falsifiable differentiators |
| owner_ack | Human ack before Pages push (operator step) |
| takedown_path | ≤1h removal path documented for the surface |
Any fail → do not publish. This page itself is drafted under that gate.