Muhammad Kashif

Engineering notes / Muhammad Kashif

Claude Code Channels, Cowork, and Dispatch: Routing Work Like Ops

Claude Code channels, Cowork, and dispatch patterns started to look like ops more than chat: queen/worker agent fleets, shared memory, routing simple tasks to cheap models, claims of deep token savings and higher effecti…

Claude Code Channels, Cowork, and Dispatch: Routing Work Like Ops

Claude Code Channels, Cowork, and Dispatch: Routing Work Like Ops

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

Claude Code channels, Cowork, and dispatch patterns started to look like ops more than chat: queen/worker agent fleets, shared memory, routing simple tasks to cheap models, claims of deep token savings and higher effective limits. Fable-class models writing trading bots from plain English and screenshot-to-UI flows sat beside self-improving loops that plan into my calendar. I kept the routing. I put the calendar on a short leash.

RooFlow-shaped fleets

Sixty-plus agents inside Claude Code with queen/worker roles, shared memory, and cheap-model routing — the Ruflo/RooFlow idea in the notes — is an ops topology. Claims around seventy-five percent token reduction and higher effective limits are hypotheses I test on my workload, not slogans I reprint in proposals.

I start with five workers, one queen, explicit job types, and a shared memory that cannot hold secrets. Fleet size is a privilege earned by clean merges.

Channels as queues

Channels beat omnipotent threads. Research channel, coding channel, calendar channel — each with tools appropriate to the risk. Cowork sits on the desktop channel with folder scopes. Dispatch means a human or queen agent assigns, not every worker grabbing the loudest task.

Queue depth is visible. Invisible backlogs are how surprises ship on Friday.

Fable-class no-code demos

Trading bots from plain-English strategy with no profit promise, screenshot-to-working UI — fine for education. I watermark financial demos and ban live capital. Screenshot-to-UI goes to staging with synthetic data.

Self-improving loops that rewrite their own goals without a human check are science fiction I do not enable on client calendars.

Daily auto-planning

Agents that plan into the calendar are useful if they only create holds labelled draft. They may not move client meetings or send invites externally without approval. My calendar is part of my reputation.

Morning briefing agents summarise; they do not silently reschedule my life to optimise a metric I did not choose.

Token routing as ops policy

Simple tasks to cheap models, hard tasks to frontier, cache shared context — the same discipline as tokenmaxxing fixes, now expressed as channel defaults. Workers inherit the default of their channel. Escalation requires a reason code.

Finance likes reason codes. So do I when the bill arrives.

Shared memory hygiene

Shared memory across workers is how fleets learn — and how one poisoned summary infects everyone. I expire memory, redact secrets, and require sources on stored claims. Dreaming-style overnight cleanup is welcome when logged.

Memory without provenance is rumour infrastructure.

What I sell

Channel topology design, queen/worker pilots with five workers, calendar agents that only draft, and token routing policies with reason codes. I do not sell sixty agents on day one with live trading attached.

Dispatch rule

Humans dispatch high risk. Queens dispatch low risk inside policy. Workers do not freelance. Cowork does not outrank the rule file because it has a mouse.

Worker freelance ban

Workers that grab tasks outside their channel get their tools reduced for a week. Queen/worker topology fails when everyone is a freelancer. Cowork mouse access does not outrank the dispatch file.

I celebrate quiet queues more than busy agents. Busy is easy. Correct routing is the craft.

Calendar draft label

Auto-planned events must include the word DRAFT in the title. External invites stay human. Channels and Cowork can suggest; reputation stays mine. Token routing does not apply to invite sends — those are always high tier, high friction.

Memory expiry job

Shared fleet memory expires on a cron with a log of what was dropped. Ruflo-style savings do not justify immortal rumour storage. Fable demos that need long memory get a dedicated, reviewed store — not the default soup.

Five workers then stop

Pilots start at five workers and one queen regardless of sixty-agent marketing. Expansion requires two weeks of clean merges and a token report finance will read. Channels beat sprawl. Cowork does not get to hire freelancers in the background.

Dispatch reviews happen Fridays. Workers that freelance lose tools Mondays. The rhythm is the control.

Trading demo watermark

Any Fable-class trading-bot demo carries a giant NO LIVE CAPITAL watermark in the UI and the README. Channels can teach. They cannot quietly connect a brokerage. Dispatch rules treat capital as always high risk.