1 / 14
Case study

Overseer of the Overseer

Sixteen projects in four days with a routed crew of language-model agents

July 2026
The problem

One AI session hits three walls.

The idea

The foreman never swings a hammer.

Architecture

One brain, many hands, two fuel tanks.

System architecture: human, orchestrator, two worker accounts with subagents, and shared services HUMAN (phone: approve / deny) single control point ORCHESTRATOR (one session, small container) ACCOUNT A WORKER 1 WORKER 2 H S O F S O F H subagents, tier per rubric: H Haiku 4.5, S Sonnet 5, O Opus 4.8, F Fable 5 ACCOUNT B WORKER 1 WORKER 2 O F S H H S O F SHARED SERVICES (available to orchestrator + all workers) RAG INDEX STATUS LOG APPROVAL GATE WATCHERS
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.

The delegate loop: brief, dispatch, work, milestone, independent verify, then pass or fail BRIEF DISPATCH right-sized model WORK MILESTONE appended to log verifier is never the builder VERIFY pass DONE fail: findings back to worker
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.

Routing decision tree from a new subtask to a model tier NEW SUBTASK MECHANICAL? yes HAIKU 4.5 no top-tier on a two-command chore no BOUNDED CHORE? yes SONNET 5 no NEEDS JUDGMENT? yes OPUS 4.8 no MULTI-HOUR BUILD? yes FABLE 5 no bottom-tier on an architecture 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.
Retrieval

Look it up before you read it.

The three-layer retrieval stack: recall, navigate, record QUERY RECALL RAG over docs, infrastructure notes, project history query before reading any file miss to logged READ ONE FILE (fallback, on miss only) NAVIGATE code graph and symbol tools for live code work only RECORD memory graph and per-topic state files decisions survive the session
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.
The gate

No answer means no.

The approval gate and the session watcher, two control-flow panels APPROVAL GATE AGENT PROPOSE (CLI) PHONE PUSH APPROVE DENY BLOCKED WAIT EXIT 0, APPROVED EXIT 1, DENIED EXIT 2, TIMEOUT timeout treated as NOT approved SESSION WATCHER WORKER SESSION observed by WATCHER (DETACHED) on cap: URGENT PUSH RESUME FILE WRITTEN SAME SESSION CONTINUES AFTER RESET never restart
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.
The campaign

Sixteen projects, four categories, one person mostly away from a keyboard.

Overnight

The night shift.

Overnight swimlane of real session timestamps across two accounts, July 10 to 11, 2026 22:00 23:00 00:00 01:00 02:00 03:00 04:00 05:00 06:00 07:00 ACCT A ACCT B COURSE SITE 22:02-22:34 USAGE DASHBOARD 00:22-01:04 COMPLIANCE PAGES 04:13-04:18 RESEARCH PLAN 04:42-05:05 BUDGET UPGRADE 22:03-22:33 FRAMEWORK EVAL 00:27-00:49 UI FEATURE 01:57-07:28, verified 07:28 proposal timed out, human asleep approved 07:19 active work blocked, waiting on approval human approved
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.
Results

What it measured.

Results panel: projects shipped by category, peak concurrency, and the campaign window PROJECTS BY CATEGORY CLIENT / REVENUE 4 INFRA / OPS 5 FAMILY ASSISTANT 2 PERSONAL ADMIN 2 plus 3 more shipped overnight into July 11, two more in flight PEAK CONCURRENCY 18 LIVE SESSIONS AT ONCE ACCT A 8 ACCT B 10 CAMPAIGN WINDOW 4 DAYS JUL 8-11, 2026 8 9 10 11
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.
What broke

Five failures, five design changes.

Discussion

When it works, when it breaks.

Works

  • Many parallelizable, medium-size projects
  • A human who is reachable
  • A budget that is metered

Breaks

  • Gate latency overnight
  • Verification overhead
  • Orchestrator context compaction
  • One human's phone as a dependency
Takeaways

Five ideas to remember.