DATA CONCLAVE · SEATTLE TECH WEEK 2026 · GENERATIVE AI poster exhibition · jul 30 2026

Operating an 11-Agent
Personal AI Staff:

Orchestration Patterns, Local LLMs, and Hard-Won Lessons

Sanket Gautam Independent · personal project, in production since Feb 2026 sanketgautam.dev
11 + 1
persona agents + one human
878
live sessions (jul 2026 audit)
82
scheduled jobs, 60–240 min heartbeats
2 / 11
seats on fully-local models (3 mlx servers)
$0
marginal cost per local token
01

One coordinator, ten specialists, enforced by config

LEGEND
frontier api seat
local mlx seat
ships drafts: human approves
SANKET
the human
COZMO
the coordinator · id: main
intake → do-or-delegate → track → roll up
claude opus 4.8 · gpt-5.5 fallback
spawns: depth ≤ 1 · ≤ 3 children
completion via sessions_yield
Coach
performance coach: cadence, not commands
opus 4.8 · #morning-coffee
Tech Lead
reviews code, raises the bar
gpt-5.5 · #build-and-experiment
Research Lead
research & strategy, end-to-end
opus 4.8 · #learning
Comms
direct comms: anything with a human recipient
opus 4.8 · #people-and-relationships
Content
content & publishing, for audiences
opus 4.8 · #content-and-social
Health
owns the body data end-to-end
opus 4.8 · #health-and-fitness
Finance
executive & finance: owns the money
opus 4.8 · #finances
Life Ops
operations & admin: keeps the machine running
opus 4.8 · #life-ops
Local Junior
local junior engineer · qwen3.6-35b moe · :8082
Local Spare
local seat (benched) · gemma-4-26b · :8801
one always-on mac studio m4 max · 128 gb · openclaw gateway under launchd · mlx_lm servers
watchdogs: 2-min health kickstart · resource monitor ram bands 70% / 80% · delegation card posted in each specialist's channel

02

Orchestration patterns
The org chart is config, not convention.

Only the coordinator holds a delegation allowlist; every other agent's is empty. Specialists cannot spawn or command each other, so the hub shape is structurally enforced.

main   allowAgents: [8 agents]
others allowAgents: []
maxSpawnDepth:  1
maxChildren:    3
completion:     sessions_yield
Delegation is a written contract.

Each handoff posts a delegation card (task-id, spec, done-criteria, deadline) in the specialist's public Slack channel, then spawns the worker. Coordination stays legible to the human.

Push, don't poll.

Children report completion via sessions_yield; polling loops are prohibited. The parent owns verification and exactly one visible completion; late duplicates are suppressed.

A staggered pulse, not a stream.

Every agent heartbeats on its own cadence (60–240 min) into its own channel: proactive, but rate-limited by design.

03

Local LLMs on MLX

Two staffed seats (plus one spare server) run entirely on the desk: the $0-marginal-cost experiment.

Qwen3.6-35B-A3B 4bit DWQMoE · ~14 GB weights · distilled + fast:8082
Local Junior
Gemma-4-26B-a4b 4bitMoE · measured ~89 tok/s on-device:8801
Local Spare
Gemma-4-31B OptiQ 4bitstaged spare capacity:8802
Division of labor.

Local takes high-volume, low-stakes work: heartbeat personas, bulk grunt tasks, framework experiments. Frontier (Opus 4.8 primary, GPT-5.5 fallback, Sonnet 5 for sub-agents) keeps judgment, delegation, and anything a human will read.

Design target: ~$100 / month

for the full staff (Feb 2026 estimate); the local seats exist to bend that curve toward zero.

04

Hard-won lessons
  1. Agent memory goes stale faster than you think.

    Twice in one week the assistant trusted its own memory file over the airline's confirmation email, and reported the wrong return date. The standing rule now: primary sources beat memory.

  2. Never present stale as fresh.

    Every claim carries its timestamp; data over 24 h old must declare its age; health reports re-fetch biometrics before speaking. "Stale data presented as current is a trust violation."

  3. Self-healing is layered supervision, not magic.

    One audit window logged 200+ gateway memory-pressure warnings. The response: a resource monitor with 70% / 80% alert bands, launchd-supervised restarts, and a 2-minute watchdog that revives the gateway itself.

  4. Escalation is a feature.

    Outbound to humans ships as a draft, never verified-and-sent in the same turn; money needs an explicit "do it." The human in the loop is a coordination mechanism, the ultimate consistency resolver.

The models were the easy part. Operating them is the system: supervision, freshness, and human gates.
built on openclaw · apple mlx · whatsapp + slack · all figures from the july 2026 system audit and live configuration · personal project, not affiliated with any employer
seat names are role archetypes, not characters · not affiliated with, endorsed by, or derived from any show, network, or studio
QR code linking to sanketgautam.dev/ensemble sanketgautam.dev/ensemble