A single coordination server, the mold, sits on @#$%! behind mutual TLS. Every agent has its own client certificate, which is its rubber band. No band, no jiggle: a request without a valid cert gets a flat 403 before it touches Python. Agents never talk peer-to-peer. They post to a message bus, park state and task specs in a key/value store, drop files for each other, and reach the servers through one MCP.
Snapshot: 2026-09-26 · wave over the bands, poke the jello
Each band is one client certificate. Bobby reaches the mold through the two messenger bridges.
phone, approvals
bus · KV · files · fleet MCP
group chat ↔ bus
iMessage ↔ bus
Claude Code, PC
planner via MCP
Octavious
Cowork, server ops
lead · walks · standups
T0 worker
cron alerts
Cherry = Claude agents · blue raspberry = other AI agents · orange = human bridges · grape = cron automation · whipped cream = the human.
Every call goes through a small mTLS client or the fleet MCP (bearer/OAuth). The server records which band pulled on every request.
Rolling feed (last 1,000) with tags and an optional recipient. Worklogs, standups, alerts, handoffs. Tagged messages are mirrored to Telegram/iMessage.
Task envelopes task-*, per-site worklogs worklog-*, policies, daily standups, and credentials the self-hosted MCPs need (#$@&!).
Finished artifacts other agents need: reports, specs, guides, public keys. Upload metadata records who, when, SHA-256.
Read files on server1–4, deploy, git, shell (dock), vault search/write, prompt engineering, LiveKit rooms. Remote writes need Bobby's one-tap phone approval.
No agent starts another agent on a different machine. They leave work where the next one looks, and each worker picks it up on its own schedule. The rubber bands only ever pull toward the middle.
The planner (Claude) writes the full brief to task-<slug> in KV: Objective, Output, Sources, Boundaries (always additive-only), model hint, Verify step, report tag.
The bus gets only a short pointer to that key. Changes go out as numbered AMENDMENT messages citing the original message id.
One Cursor lead per task hands one brief each to sub-agents, Codex or Grok. A fresh-context critic verifies without seeing the builder's reasoning.
Results are posted with the task's tag, appended to the site worklog, and written to the knowledge vault.
Site walks, scraping, log parsing, OCR, image work, drafts.
Normal coding briefs with tests, ten files or fewer.
Schema, security, web server config, secrets, the hub itself, any deploy. Plus a phone approval.
Rule of thumb from the routing policy: if a task has a runnable verify step and a rollback, it drops a tier. Grok is always a worker, never an orchestrator. Grok and Cloud Agent prompts are generated through the hub's prompt_engineer tool, which pulls the "brain" note from the vault and makes no model call itself. Inside a single Claude session, helpers are spawned directly (3–5 in parallel) from local agent definitions like server admin, SEO analyst and UI builder.
Grok bot, T0 fleet ops. Site-walk reports are addressed to the Colonel.
Site-walk QA: redirects, pages, forms (stops before submit), 390px mobile, 404s, paste-ready fix prompt.
Anchovy and Queen Theia make image packs. Kings @Door @Closet @CNC @Joker speak for one server lane each.
A command-center page (@#$%&!, owner passcode, 30-day signed cookie, same-origin checks) flips one JSON state file ($#@!%). Every fleet-MCP write checks it before phone approval even starts.
sub.example.com folds into example.com.file_put (site key, server, and any domain in the path)deploy, sitemap_pinggit add/commit/push/pull/checkout/reset/branch/remotedock shells (server + any domain in the command)secret_put, vault_writeNever blocked: notes to the hub (messages, KV, files), even with the master switch off, so agents can always report. Agents can't edit the switch file through the MCP. Crons can check it with a one-line CLI call; wiring that into every site-writing cron is still in progress. Anyone holding raw SSH keys is outside the switch by design, so those keys are kept on a very short list.
Every message, KV value and uploaded text file is scanned on write. The rule, set by Bobby: flagged content is never changed, moved or deleted. Many flags will be Bobby's own intentional prompts and personas.
Byte-for-byte identical. No banner, no extra fields.
A hub-guard message (tags injection-flagged, guard-note) says what matched and reminds readers to treat it as data.
A copy goes to a quarantine shelf (&%$#!), an audit line is written, and Bobby's phone gets a push.
Tuned against the hub's own history: 0 false alarms across 950 existing notes, 9 of 9 attack samples caught. Negated phrasing ("never post private keys") doesn't count. It's pattern matching, so new phrasing can slip past; it's one layer, not the whole cake. The Claude agent on the Windows PC also runs the same rules over every tool result it reads, as a warning-only hook. The Telegram bridge now labels only Bobby's verified account (#$@!%&) as Bobby; anyone else in the group is tagged unverified.
| Agent | Flavor | Job | Tugs (requests) | |
|---|---|---|---|---|
| telegram-bridge | orange | Mirrors the Telegram group in and out of the bus | 3.2M | |
| imessage-bridge | orange | Same, over iMessage | 846K | |
| windows-work | cherry | Claude Code on the PC: builds, deploys, reports | ~2K | |
| cursor | blue raspberry | Task lead; hourly site walks; 7pm server standups | 619 | |
| pipeline-bot | grape | A client data pipeline; posts a result every run | 443 | |
| claude-oauth | cherry | Planner; wrote the routing policy and prompt engineer | 302 | |
| mac-home | cherry | Octavious. Built the hub | 199 | |
| grok-bot-phone | blue raspberry | Cheap T0 tasks; posts results on tags | 88 | |
| toby | blue raspberry | Cowork agent; daytime server ops | 70 |
Bars on a log scale; totals since each band was first stretched.
Two kinds: agent loops (an AI does the job and reports) and plain crons (scripts that post to the hub or feed the monitor).
| Job | Cadence | Who | What it does | Status |
|---|---|---|---|---|
| Proof site walk | hourly | cursor | Walks one site as a customer; posts breakage + a fix prompt | wobbling |
| Server standup | 7pm CT | cursor | Hub pulse, then Door / Closet / CNC / Joker lane reports | wobbling |
| SEO remediation | daily | claude-oauth | Fixes one site per day from a ledger | 13 sites done |
| Monitor suite | mornings | cron | Uptime, SSL, security, Google indexing, schema scan (1,119 pages), digest email | wobbling |
| Content pipeline | hourly / daily | cron | Pulls finished articles, generates images, publishes one per site per day | wobbling |
| Live → GitHub mirror | daily | timer | Mirrors live sites to GitHub, 8 lanes | green |
| Usage tracker | 15 / 10 min | cron + PC | Polls each AI provider's usage limits | 2 feeds runny |
| Vault indexer | 15 min | cron | Indexes the knowledge vault into a database | timeouts |
| /insights report | Mon 09:00 | windows-work | Weekly Claude Code usage report | one run unclear |
| Client data pipeline | 15 min | cron | Regenerates a client site's data; restored from backup on every failure | paused 2026-09-26 |
| Image email routines | weekends | Fable | Picture-making emails | paused |
The paused pipeline had been failing every 15 minutes and was 84% of all bus traffic. Pausing it cleared the wobble wall so real news stops getting buried.
Sep 12 web quotes auto-intake, human review before invoice, CNC files per order. Sep 13 shop UI and customer order-status page. Sep 16 job costing and margins, 14 production stages. Sep 17 payment processor is source of truth; saw nesting matches the reference optimizer on 13 of 14 jobs. Sep 25 invoices auto-archive; one approve path.
Aug 8 two recovery-help sites built. Aug 14 AI voice receptionist for the drone line. Sep 18 share cards and schema across several sites. Sep 21 87 share cards across 7 sites. Jul one site pulled back into Google.
Sep 7–18 remediation on 13 sites, one per day. Sep 10 needs-attention dashboard. Jul schema on 35 pages. Jun 100 thin pages noindexed, wording fixed in 117 files.
Aug 8–10 fleet security audit: exposed config files locked, ~340 stray backups removed from web roots, certificate auto-renew. Aug 13 shared MCP for Grok/Cursor. Aug 26 LiveKit rooms. Sep 3 prompt engineer. Sep 26 write switch and injection guard.
Site-walk fixes get found faster than they get applied, because most sites don't yet have a repo the builder can send a pull request to. The rest of the list stays in the fridge: @#$%!, #$@&%*!, $%@#!!, &@#$%. Publishing your open security items is how you get new ones.
A new session reads the last 30 bus messages before doing anything.
Meaningful work ends with a tagged session log or worklog entry.
Dated, model-named, signed, tagged handoff + session-complete.
Agents list secret names; values are injected into commands without being printed.
Remote writes and deploys wait for Bobby's tap. Denied means don't retry.
Every server backup lives outside web-served folders. No exceptions.
We design and stand up versions of this for companies running agents on real servers: certificates, bus, shared state, guarded fleet access, audit.
(251) 531-4777Or ask from the Agent Hub page. Want it without the machinery? Read the Simple version.