All thoughts

Agent OS, Part 3: Your own system

Stages 5 and 6: personal agent harnesses, then fast System 1 judgment next to slow LLM reasoning.

How it works

Day ~265 · Sep 2026 Draw your own harness

Day ~265: my own harness, with a human gateThe agent deck now holds several agents. A new platform above Eval holds a single key: the human approval gate, lit in amber.diagnose → fixMemory3 pillars + gateAgentsseveral agentsTracewhat ranEvalhealthy? good?Human gateapprovals

Personal agents are the same boxes in different shells: VM or compute, connectors, routines, approvals, memory. Grok Bot is the multi-agent desktop I actually run.

Build products for agents (APIs, MCP) as much as for people.

Day ~270 · Sep 2026 Put System 1 next to System 2

Day ~270: a fast System 1 check before the LLMThe same stack. A small cube now sits beside the eval gate, lit in amber: a cheap, fast classifier that decides first.diagnose → fixMemory3 pillars + gateAgentsseveral agentsTracewhat ranEvalSystem 1 firstHuman gateapprovals

The terms come from Daniel Kahneman: System 1 is fast and intuitive, System 2 is slow and deliberate. In an agent, System 1 is fast, structured judgment (small classifiers and rules). System 2 is slow LLM reasoning. Use System 1 for triage, routing and eval gates. Use System 2 for deep work.

My take

What I learnedCool demos aren't use cases, and portable memory beats locked silos.

What I learnedFast judgment enhances slow reasoning. It doesn't replace it.

Try this

  1. Draw your own harness once. Name what's shared and what's per-user. Name what needs a human gate.
  2. Before every LLM call in a loop, ask whether a cheap classifier should decide first.

Sources