Sixteen projects in four days with a routed crew of language-model agents
July 2026
The problem
One AI session hits three walls.
Context. A day of real work across a dozen projects does not fit in one window.
Cost. Paying top-tier rates for mechanical work wastes a metered budget.
Trust. An agent's report about its own work is not evidence.
The idea
The foreman never swings a hammer.
The orchestrator plans, dispatches, and verifies. It never builds.
Every worker carries the same rule for its own project.
The human oversees the overseer, from a phone.
Architecture
One brain, many hands, two fuel tanks.
Figure 1. One orchestrator directs two worker accounts, each fanning out to tier-routed subagents, with shared services on the side and a single human approval point at the top.
The delegate loop
The verifier is never the builder.
Figure 2. Each unit of work moves through dispatch, execution, and a logged milestone, then an independent verifier passes it to done or sends it back with findings.
Routing
Right-size every model call.
Figure 3. A new subtask falls through mechanical, bounded, judgment, and build questions until it lands on the cheapest model that clears the bar.
Do not spend a Fable agent on a two-command chore.
Do not send Haiku at an architectural decision.
Retrieval
Look it up before you read it.
Figure 4. A query hits recall first and only falls through to a single file read on a miss, while navigate and record serve code work and durable state on their own tracks.
Recall. Query the index before opening any file.
Navigate. Trace code with a graph, not a grep.
Record. Durable decisions survive a context reset.
The gate
No answer means no.
Figure 5. The gate blocks on a phone tap and treats a timeout as a denial, while a detached watcher resumes the same session after a usage cap instead of restarting it.
Exit 0 is approved. Exit 1 is denied. Exit 2 is timeout.
Timeout is treated as not approved.
The campaign
Sixteen projects, four categories, one person mostly away from a keyboard.
A client's course website, built from her own voice memos.
An infrastructure audit that found and rotated an exposed key.
A family member's assistant, with a budgeting upgrade of its own.
20,367 promotional emails archived.
Overnight
The night shift.
Figure 6. Real timestamps from the shared status log show two accounts running work in parallel through the night, with the one overnight approval gate visible as a stalled bar that resolves at the morning tap.
Two accounts ran work in parallel through the night.
Two deploy proposals timed out overnight and waited for the morning tap.
Results
What it measured.
Figure 7. Thirteen projects at the mid-run record, with three more finished overnight, spanned client, infrastructure, family, and personal work; concurrency peaked at eighteen live sessions across the four-day window.
Peak concurrency: eighteen live sessions at once, eight on one account and ten on the other.
What broke
Five failures, five design changes.
Phantom rogue orchestrator. A monitoring agent invented a threat that did not exist.
Dead email connector. It worked interactively and was silently dead headless.
The em-dash. One character broke the approval push, the system's single control point.
Stale scrollback. False cap reads, then a real 429 outage under fast polling.
Invisible deliverables. Second-account artifacts rendered for nobody.
Paper cuts. A sixth bucket of small mechanical failures, logged in the paper.