Skip to content

Knoxville, TN ET Director of AI

Case study March to November 2026

Three generations of one machine

I build AI systems, measure whether they earn their keep, keep what does and replace what doesn't.

  1. Generation 1

    Lobster Actual and Hazel

    March to September 2026

  2. The decision

    Decommissioned both, verified with a cold boot

    July to September 2026

  3. Generation 2

    Inference and utility hub

    September 2026, now the base layer

  4. Generation 3 // Current

    Orchestrated Claude Code

    November 2026, current

Noel DeLisle Director of AI, GO2 Partners

Since March 2026 one machine has run three generations of AI systems. Each one was built to do real work and judged by whether it did. The rule is the same one I apply at GO2: governance is enforced in the system, not in a policy deck.

Generation 1 // March to September 2026

Lobster Actual and Hazel

Lobster Actual was a self-hosted 10-agent system built on OpenClaw. It ran as a hardened system service, with agents for orchestration, code review, research and intelligence, publishing and analysis.

  • Models. A local-to-cloud fallback chain. Claude was structurally unreachable from sub-agent routing, so cost control never depended on an agent following a rule.
  • Governance in three layers. Per-agent identity, a canonical rule set and a runtime contract.
  • Observability. LobsterOps logged agent events, with Grafana and Prometheus for monitoring.
  • Delivery. Autonomous review-then-merge pipelines across six product repos.

Hazel was a separate personal-life agent built on Hermes Agent and isolated from Lobster Actual by a strict credential wall. Two agents, one machine, zero shared credentials.

I wrote about the first build in I Built a $2,800 Mac mini Into a 24/7 Autonomous AI Agent. Here's Everything That Went Wrong (and Right). [1]

The decision // July to September 2026

The decision

By July both agents had gone largely unused for weeks. I treated that as data. The question was whether the stack still earned its place, and the answer was no. I decommissioned both in September 2026 and verified the shutdown with a cold boot.

Most AI programs skip this step. Usage drops, the system keeps running and nobody owns the call to turn it off.

Retiring a system on evidence is part of governing it.

Generation 2 // September 2026, now the base layer

Inference and utility hub

The second build started from usage, not ambition.

  • A hybrid local and cloud roster of 11 models served across my private network.
  • An MLX server measured at about 2x Ollama throughput on the same 27B model: about 14.5 versus 7.7 tokens per second. [2]
MLXabout 14.5 tok/s
Ollamaabout 7.7 tok/s
Throughput on the same 27B model, same machine. Source: my own measurement [2]
  • A phone-friendly chat interface and uptime monitoring.
  • Encrypted offsite backups with a verified restore.

The standing rule: run it for a month and let real usage decide what gets added.

Generation 3 // November 2026, current

Orchestrated Claude Code

Generation 3 builds on Generation 2 instead of replacing it. The inference hub earned its place and stays as the base layer. Orchestration sits on top.

  • An orchestrator agent runs role-scoped bots for engineering, experiments, ideas and design. Claude Code on the same machine does all the building.
  • A model and effort routing rubric decides which model handles which work, with an escalation policy when a step fails.
  • Credential and connector isolation is enforced in software, so work data and keys cannot reach agent sessions.
Credential isolation by botIllustrative
  • EngineeringRead Product repos, open issuesWrite Feature branches, pull requestsNeeds approval Merge to main
  • ExperimentsRead Sandbox repo, public docsWrite Sandbox branchesNeeds approval Any new connector
  • IdeasRead Notes, backlogWrite Backlog entriesNeeds approval None. No code access.
  • DesignRead Reference library, design canvasWrite Design files, design branchesNeeds approval Publishing outside the canvas
  • OrchestratorRead Task queue, run logsWrite Task assignmentsNeeds approval Model escalation, scope changes

Keys and work connectors are stripped in software before any session starts. Scope limits are set per bot.

  • Every change ships through a branch, a pull request and automated Claude review.
Example runMergedDesign bot: build reference board
  1. 09:02:11 Orchestrator2sAssigns the task to the Design bot.
  2. 09:02:13 Claude Code6m 12sCreates a branch and builds.
  3. 09:08:25 Pull request4sOpens with a summary of the change.
  4. 09:08:29 Claude review1m 08sAutomated review. 2 comments, 0 blocking.
  5. 09:09:37 Human merge gate3m 41sWaits for a person. Nothing merges without approval. Approved. Merged to main.

5 steps 1 human approval 11m 07sIllustrative. Sample task and times.

  • A design-first pipeline: reference libraries, a design canvas, taste and component skills and an accessibility and UX audit.

Three generations // One rule

What carries over

Three generations, one rule: keep what earns its place, replace what doesn't and put the rules inside the system. At GO2 that means a formal intake process for every AI tool and a policy that sets how much autonomy agents get. On my own hardware it means per-agent identity, credential walls and a review gate on every change.

The thinking behind both is in Agents Don't Need More Power. They Need Better Gates., Human in the Loop Is a Design Claim and The Rogue Agent Problem Is Already Here.

Sources

  1. [1] I Built a $2,800 Mac mini Into a 24/7 Autonomous AI Agent. Here's Everything That Went Wrong (and Right). Essay, March 22, 2026.
  2. [2] My own measurement: the same 27B model on the same machine, about 14.5 versus 7.7 tokens per second.