← Long Nguyen
Product / AI Systems · Atomic Projects · 2025–2026

Standing up an AI agent operating system

Teaching an AI agent to run the parts of the business a founder didn't have time for.

n8nClaudeNotionAPI Design

The problem

Ben needed to scale fast, and he'd already bet on AI as part of how he'd do it. I'd previously helped him clean up his Notion workspace — clear divisions, clear SOPs, clear documentation. What he still lacked wasn't information, it was a system to actually scale on top of, across every function of the business at once.

What I did

A secure n8n portal handles tool access, n8n itself does orchestration, and Claude sits underneath doing the reasoning — but ‘an agent that calls tools’ isn't an operating system on its own. It needed short-term memory for the task at hand, long-term memory so it doesn't relearn the business every run, a registry of tools it's actually allowed to touch, a guardrail that stops it before anything risky, and logs that feed back into what it knows next time:

Trigger: Slack messageSlack messageTrigger: New CRM leadNew CRM leadTrigger: Scheduled checkScheduled checkTrigger: Inbound emailInbound emailMEMORYSHORT-TERMActive task context,conversation stateLong-term: process docs, SOPs, rulesLONG-TERMProcess docs, SOPs,decision rules(vector store)↓ updated by every logged runAGENT COREClaude — reasoning loopPlanPlanActActObserveObserveReflectReflectrepeats until the task is done or it asks for helpTOOL REGISTRYTool call: CRMCRMTool call: SlackSlackTool call: InvoicingInvoicingTool call: CalendarCalendarTool call: DocsDocsTool call: EmailEmailHuman approval required before high-stakes actionsGUARDRAILSHuman approval required — payments, external messages, record deletionAction: Write to CRMWrite to CRMAction: Send Slack updateSend Slack updateAction: Update docs & SOPsUpdate docs & SOPsAction: Escalate to humanEscalate to humanMONITORING & LOGSEvery step logged — alerts to Slack on failure, feeds back into long-term memory
The plan → act → observe → reflect loop repeats until the task is done or the agent asks for a human.

On top of that stack, we designed a discovery → documentation → scale framework. Discovery: the agent calls internal tools directly via API, guided by a file telling it exactly where to read documentation from. Documentation & planning: Claude reads the process, understands it, and builds a plan. Scale: I layer in automation that requires standardization, with Claude monitoring and doing the parts of the work that need active input.

1. DISCOVERY: Agent calls tools directly — guided by a docs-location file1. DISCOVERYAgent calls tools directlyguided by a docs-location file25% agent-run2. DOCUMENT & PLAN: Claude maps the process — reads it, then drafts a plan2. DOCUMENT & PLANClaude maps the processreads it, then drafts a plan55% agent-run3. SCALE: Monitored automation — Claude runs the standardized parts3. SCALEMonitored automationClaude runs the standardized parts85% agent-run

What happened

Ben now directs the business and Claude executes across it — discovery, documentation, and monitored automation all running on the same framework instead of five separate ad-hoc efforts.

A short screen-recorded walkthrough of the system in action.
10× project volume handled with zero new hires; the business now runs semi-autonomously on the system
← Back to all projects